mirror of
https://github.com/Sygil-Dev/sygil-webui.git
synced 2024-12-14 05:58:18 +03:00
Merge branch 'dev' into master
This commit is contained in:
cmdr2
GitHub
commit
5e725e036e
@ -21,7 +21,7 @@ Come to our [Discord Server](https://discord.gg/gyXNe4NySY) or use [Discussions]
|
||||
|
||||
Check the [Contribution Guide](CONTRIBUTING.md)
|
||||
|
||||
[sygil-dev](https://github.com/sygil-dev) main devs:
|
||||
[Sygil-Dev](https://github.com/Sygil-Dev) main devs:
|
||||
|
||||
*  [hlky](https://github.com/hlky)
|
||||
* [ZeroCool940711](https://github.com/ZeroCool940711)
|
||||
|
||||
@ -129,7 +129,7 @@
|
||||
"\n",
|
||||
"## Streamlit\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"**Features:**\n",
|
||||
"\n",
|
||||
@ -148,7 +148,7 @@
|
||||
"\n",
|
||||
"## Gradio\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"**Features:**\n",
|
||||
"\n",
|
||||
@ -166,7 +166,7 @@
|
||||
"\n",
|
||||
"### GFPGAN\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"Lets you improve faces in pictures using the GFPGAN model. There is a checkbox in every tab to use GFPGAN at 100%, and also a separate tab that just allows you to use GFPGAN on any picture, with a slider that controls how strong the effect is.\n",
|
||||
"\n",
|
||||
@ -176,7 +176,7 @@
|
||||
"\n",
|
||||
"### RealESRGAN\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"Lets you double the resolution of generated images. There is a checkbox in every tab to use RealESRGAN, and you can choose between the regular upscaler and the anime version.\n",
|
||||
"There is also a separate tab for using RealESRGAN on any picture.\n",
|
||||
@ -265,7 +265,56 @@
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"source": [
|
||||
"# Setup"
|
||||
"# Config options for Colab instance\n",
|
||||
"> Before running, make sure GPU backend is enabled. (Unless you plan on generating with Stable Horde)\n",
|
||||
">> Runtime -> Change runtime type -> Hardware Accelerator -> GPU (Make sure to save)"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "iegma7yteERV"
|
||||
}
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"#@markdown WebUI repo (and branch)\n",
|
||||
"repo_name = \"Sygil-Dev/sygil-webui\" #@param {type:\"string\"}\n",
|
||||
"repo_branch = \"dev\" #@param {type:\"string\"}\n",
|
||||
"\n",
|
||||
"#@markdown Mount Google Drive\n",
|
||||
"mount_google_drive = True #@param {type:\"boolean\"}\n",
|
||||
"save_outputs_to_drive = True #@param {type:\"boolean\"}\n",
|
||||
"#@markdown Folder in Google Drive to search for custom models\n",
|
||||
"MODEL_DIR = \"\" #@param {type:\"string\"}\n",
|
||||
"\n",
|
||||
"#@markdown Enter auth token from Huggingface.co\n",
|
||||
"#@markdown >(required for downloading stable diffusion model.)\n",
|
||||
"HF_TOKEN = \"\" #@param {type:\"string\"}\n",
|
||||
"\n",
|
||||
"#@markdown Select which models to prefetch\n",
|
||||
"STABLE_DIFFUSION = True #@param {type:\"boolean\"}\n",
|
||||
"WAIFU_DIFFUSION = False #@param {type:\"boolean\"}\n",
|
||||
"TRINART_SD = False #@param {type:\"boolean\"}\n",
|
||||
"SD_WD_LD_TRINART_MERGED = False #@param {type:\"boolean\"}\n",
|
||||
"GFPGAN = True #@param {type:\"boolean\"}\n",
|
||||
"REALESRGAN = True #@param {type:\"boolean\"}\n",
|
||||
"LDSR = True #@param {type:\"boolean\"}\n",
|
||||
"BLIP_MODEL = False #@param {type:\"boolean\"}\n",
|
||||
"\n"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "OXn96M9deVtF"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"source": [
|
||||
"# Setup\n",
|
||||
"\n",
|
||||
">Runtime will crash when installing conda. This is normal as we are forcing a restart of the runtime from code.\n",
|
||||
"\n",
|
||||
">Just hit \"Run All\" again. 😑"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "IZjJSr-WPNxB"
|
||||
@ -277,6 +326,7 @@
|
||||
"id": "eq0-E5mjSpmP"
|
||||
},
|
||||
"source": [
|
||||
"#@title Make sure we have access to GPU backend\n",
|
||||
"!nvidia-smi -L"
|
||||
],
|
||||
"execution_count": null,
|
||||
@ -285,14 +335,14 @@
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"#@title Install miniConda (mamba)\n",
|
||||
"!pip install condacolab\n",
|
||||
"import condacolab\n",
|
||||
"condacolab.install_from_url(\"https://github.com/conda-forge/miniforge/releases/download/4.14.0-0/Mambaforge-4.14.0-0-Linux-x86_64.sh\")\n",
|
||||
"\n",
|
||||
"import condacolab\n",
|
||||
"condacolab.check()\n",
|
||||
"\n",
|
||||
"# The runtime will crash after this, its normal as we are forcing a restart of the runtime from code. Just hit \"Run All\" again."
|
||||
"# The runtime will crash here!!! Don't panic! We planned for this remember?"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "cDu33xkdJ5mD"
|
||||
@ -303,9 +353,13 @@
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"!git clone https://github.com/Sygil-Dev/sygil-webui.git\n",
|
||||
"%cd /content/sygil-webui/\n",
|
||||
"!git checkout dev\n",
|
||||
"#@title Clone webUI repo and download font\n",
|
||||
"import os\n",
|
||||
"REPO_URL = os.path.join('https://github.com', repo_name)\n",
|
||||
"PATH_TO_REPO = os.path.join('/content', repo_name.split('/')[1])\n",
|
||||
"!git clone {REPO_URL}\n",
|
||||
"%cd {PATH_TO_REPO}\n",
|
||||
"!git checkout {repo_branch}\n",
|
||||
"!git pull\n",
|
||||
"!wget -O arial.ttf https://github.com/matomo-org/travis-scripts/blob/master/fonts/Arial.ttf?raw=true"
|
||||
],
|
||||
@ -318,7 +372,10 @@
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"!mamba install cudatoolkit=11.3 git numpy=1.22.3 pip=20.3 python=3.8.5 pytorch=1.11.0 scikit-image=0.19.2 torchvision=0.12.0 -y"
|
||||
"#@title Install dependencies\n",
|
||||
"!mamba install cudatoolkit=11.3 git numpy=1.22.3 pip=20.3 python=3.8.5 pytorch=1.11.0 scikit-image=0.19.2 torchvision=0.12.0 -y\n",
|
||||
"!python --version\n",
|
||||
"!pip install -r requirements.txt"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "dmN2igp5Yk3z"
|
||||
@ -329,52 +386,29 @@
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"#@title Install dependencies.\n",
|
||||
"!python --version\n",
|
||||
"!pip install -r requirements.txt"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "vXX0OaR8KyLQ"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"#@title Install localtunnel to openGoogle's ports\n",
|
||||
"!npm install localtunnel"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "FHyVuT5aSM2G"
|
||||
"id": "Nxaxfgo_F8Am"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"source": [
|
||||
"#Launch the WebUI"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "csi6cj6gQZmC"
|
||||
}
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"#@title Mount Google Drive\n",
|
||||
"import os\n",
|
||||
"mount_google_drive = True #@param {type:\"boolean\"}\n",
|
||||
"save_outputs_to_drive = True #@param {type:\"boolean\"}\n",
|
||||
"\n",
|
||||
"#@title Mount Google Drive (if selected)\n",
|
||||
"if mount_google_drive:\n",
|
||||
" # Mount google drive to store your outputs.\n",
|
||||
" # Mount google drive to store outputs.\n",
|
||||
" from google.colab import drive\n",
|
||||
" drive.mount('/content/drive/', force_remount=True)\n",
|
||||
"\n",
|
||||
"if save_outputs_to_drive:\n",
|
||||
" os.makedirs(\"/content/drive/MyDrive/sygil-webui/outputs\", exist_ok=True)\n",
|
||||
" os.symlink(\"/content/drive/MyDrive/sygil-webui/outputs\", \"/content/sygil-webui/outputs\", target_is_directory=True)\n"
|
||||
" # Make symlink to redirect downloads\n",
|
||||
" OUTPUT_PATH = os.path.join('/content/drive/MyDrive', repo_name.split('/')[1], 'outputs')\n",
|
||||
" os.makedirs(OUTPUT_PATH, exist_ok=True)\n",
|
||||
" os.symlink(OUTPUT_PATH, os.path.join(PATH_TO_REPO, 'outputs'), target_is_directory=True)\n"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "pcSWo9Zkzbsf"
|
||||
@ -385,20 +419,114 @@
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"#@title Enter Huggingface token\n",
|
||||
"!git config --global credential.helper store\n",
|
||||
"!huggingface-cli login"
|
||||
"#@title Pre-fetch models\n",
|
||||
"%cd {PATH_TO_REPO}\n",
|
||||
"# make list of models we want to download\n",
|
||||
"model_list = {\n",
|
||||
" 'stable_diffusion': f'{STABLE_DIFFUSION}',\n",
|
||||
" 'waifu_diffusion': f'{WAIFU_DIFFUSION}',\n",
|
||||
" 'trinart_stable_diffusion': f'{TRINART_SD}',\n",
|
||||
" 'sd_wd_ld_trinart_merged': f'{SD_WD_LD_TRINART_MERGED}',\n",
|
||||
" 'gfpgan': f'{GFPGAN}',\n",
|
||||
" 'realesrgan': f'{REALESRGAN}',\n",
|
||||
" 'ldsr': f'{LDSR}',\n",
|
||||
" 'blip_model': f'{BLIP_MODEL}'}\n",
|
||||
"download_list = {k for (k,v) in model_list.items() if v == 'True'}\n",
|
||||
"\n",
|
||||
"# get model info (file name, download link, save location)\n",
|
||||
"import yaml\n",
|
||||
"from pprint import pprint\n",
|
||||
"with open('configs/webui/webui_streamlit.yaml') as f:\n",
|
||||
" dataMap = yaml.safe_load(f)\n",
|
||||
"models = dataMap['model_manager']['models']\n",
|
||||
"\n",
|
||||
"# copy script from model manager\n",
|
||||
"import requests, time\n",
|
||||
"from requests.auth import HTTPBasicAuth\n",
|
||||
"\n",
|
||||
"def download_file(file_name, file_path, file_url):\n",
|
||||
" os.makedirs(file_path, exist_ok=True)\n",
|
||||
" if os.path.exists(os.path.join(MODEL_DIR , file_name)):\n",
|
||||
" print( file_name + \"found in Google Drive\")\n",
|
||||
" print( \"Creating symlink...\")\n",
|
||||
" os.symlink(os.path.join(MODEL_DIR , file_name), os.path.join(file_path, file_name))\n",
|
||||
" elif not os.path.exists(os.path.join(file_path , file_name)):\n",
|
||||
" print( \"Downloading \" + file_name + \"...\", end=\"\" )\n",
|
||||
" token = None\n",
|
||||
" if \"huggingface.co\" in file_url:\n",
|
||||
" token = HTTPBasicAuth('token', HF_TOKEN)\n",
|
||||
" try:\n",
|
||||
" with requests.get(file_url, auth = token, stream=True) as r:\n",
|
||||
" starttime = time.time()\n",
|
||||
" r.raise_for_status()\n",
|
||||
" with open(os.path.join(file_path, file_name), 'wb') as f:\n",
|
||||
" for chunk in r.iter_content(chunk_size=8192):\n",
|
||||
" f.write(chunk)\n",
|
||||
" if ((time.time() - starttime) % 60.0) > 2 :\n",
|
||||
" starttime = time.time()\n",
|
||||
" print( \".\", end=\"\" )\n",
|
||||
" print( \"done\" )\n",
|
||||
" print( \" \" + file_name + \" downloaded to \\'\" + file_path + \"\\'\" )\n",
|
||||
" except:\n",
|
||||
" print( \"Failed to download \" + file_name + \".\" )\n",
|
||||
" else:\n",
|
||||
" print( file_name + \" already exists.\" )\n",
|
||||
"\n",
|
||||
"# download models in list\n",
|
||||
"for model in download_list:\n",
|
||||
" model_name = models[model]['model_name']\n",
|
||||
" file_info = models[model]['files']\n",
|
||||
" for file in file_info:\n",
|
||||
" file_name = file_info[file]['file_name']\n",
|
||||
" file_url = file_info[file]['download_link']\n",
|
||||
" if 'save_location' in file_info[file]:\n",
|
||||
" file_path = file_info[file]['save_location']\n",
|
||||
" else: \n",
|
||||
" file_path = models[model]['save_location']\n",
|
||||
" download_file(file_name, file_path, file_url)\n",
|
||||
"\n",
|
||||
"# add custom models not in list\n",
|
||||
"CUSTOM_MODEL_DIR = os.path.join(PATH_TO_REPO, 'models/custom')\n",
|
||||
"if MODEL_DIR != \"\":\n",
|
||||
" MODEL_DIR = os.path.join('/content/drive/MyDrive', MODEL_DIR)\n",
|
||||
" if os.path.exists(MODEL_DIR):\n",
|
||||
" custom_models = os.listdir(MODEL_DIR)\n",
|
||||
" custom_models = [m for m in custom_models if os.path.isfile(MODEL_DIR + '/' + m)]\n",
|
||||
" os.makedirs(CUSTOM_MODEL_DIR, exist_ok=True)\n",
|
||||
" print( \"Custom model(s) found: \" )\n",
|
||||
" for m in custom_models:\n",
|
||||
" print( \" \" + m )\n",
|
||||
" os.symlink(os.path.join(MODEL_DIR , m), os.path.join(CUSTOM_MODEL_DIR, m))\n",
|
||||
"\n"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "IsbG7fvIrKwg"
|
||||
"id": "vMdmh81J70yA"
|
||||
},
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"source": [
|
||||
"# Launch the web ui server\n",
|
||||
"### (optional) JS to prevent idle timeout:\n",
|
||||
"Press 'F12' OR ('CTRL' + 'SHIFT' + 'I') OR right click on this website -> inspect. Then click on the console tab and paste in the following code.\n",
|
||||
"```js,\n",
|
||||
"function ClickConnect(){\n",
|
||||
"console.log(\"Working\");\n",
|
||||
"document.querySelector(\"colab-toolbar-button#connect\").click()\n",
|
||||
"}\n",
|
||||
"setInterval(ClickConnect,60000)\n",
|
||||
"```"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "pjIjiCuJysJI"
|
||||
}
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"#@title <-- Press play on the music player to keep the tab alive (Uses only 13MB of data)\n",
|
||||
"#@title Press play on the music player to keep the tab alive (Uses only 13MB of data)\n",
|
||||
"%%html\n",
|
||||
"<b>Press play on the music player to keep the tab alive, then start your generation below (Uses only 13MB of data)</b><br/>\n",
|
||||
"<audio src=\"https://henk.tech/colabkobold/silence.m4a\" controls>"
|
||||
@ -409,27 +537,10 @@
|
||||
"execution_count": null,
|
||||
"outputs": []
|
||||
},
|
||||
{
|
||||
"cell_type": "markdown",
|
||||
"source": [
|
||||
"JS to prevent idle timeout:\n",
|
||||
"\n",
|
||||
"Press F12 OR CTRL + SHIFT + I OR right click on this website -> inspect. Then click on the console tab and paste in the following code.\n",
|
||||
"\n",
|
||||
"function ClickConnect(){\n",
|
||||
"console.log(\"Working\");\n",
|
||||
"document.querySelector(\"colab-toolbar-button#connect\").click()\n",
|
||||
"}\n",
|
||||
"setInterval(ClickConnect,60000)"
|
||||
],
|
||||
"metadata": {
|
||||
"id": "pjIjiCuJysJI"
|
||||
}
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"source": [
|
||||
"#@title Open port 8501 and start Streamlit server. Open link in 'link.txt' file in file pane on left.\n",
|
||||
"#@title Run localtunnel and start Streamlit server. ('Ctrl' + 'left click') on link in the 'link.txt' file. (/content/link.txt)\n",
|
||||
"!npx localtunnel --port 8501 &>/content/link.txt &\n",
|
||||
"!streamlit run scripts/webui_streamlit.py --theme.base dark --server.headless true 2>&1 | tee -a /content/log.txt"
|
||||
],
|
||||
|
||||
@ -23,6 +23,7 @@ general:
|
||||
streamlit_telemetry: False
|
||||
default_theme: dark
|
||||
huggingface_token: ''
|
||||
stable_horde_api: '0000000000'
|
||||
gpu: 0
|
||||
outdir: outputs
|
||||
default_model: "Stable Diffusion v1.5"
|
||||
@ -65,6 +66,10 @@ general:
|
||||
update_preview: True
|
||||
update_preview_frequency: 10
|
||||
|
||||
admin:
|
||||
hide_server_setting: False
|
||||
hide_browser_setting: False
|
||||
|
||||
debug:
|
||||
enable_hydralit: False
|
||||
|
||||
@ -230,7 +235,8 @@ img2img:
|
||||
step: 0.01
|
||||
# 0: Keep masked area
|
||||
# 1: Regenerate only masked area
|
||||
mask_mode: 0
|
||||
mask_mode: 1
|
||||
noise_mode: "Matched Noise"
|
||||
mask_restore: False
|
||||
# 0: Just resize
|
||||
# 1: Crop and resize
|
||||
@ -315,7 +321,7 @@ gfpgan:
|
||||
strength: 100
|
||||
|
||||
textual_inversion:
|
||||
pretrained_model_name_or_path: "models/diffusers/stable-diffusion-v1-4"
|
||||
pretrained_model_name_or_path: "models/diffusers/stable-diffusion-v1-5"
|
||||
tokenizer_name: "models/clip-vit-large-patch14"
|
||||
|
||||
|
||||
|
||||
467
db.json
Normal file
467
db.json
Normal file
@ -0,0 +1,467 @@
|
||||
{
|
||||
"stable_diffusion": {
|
||||
"name": "stable_diffusion",
|
||||
"type": "ckpt",
|
||||
"description": "Generalist AI image generating model. The baseline for all finetuned models.",
|
||||
"version": "1.5",
|
||||
"style": "generalist",
|
||||
"nsfw": false,
|
||||
"download_all": true,
|
||||
"requires": [
|
||||
"clip-vit-large-patch14"
|
||||
],
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/ldm/stable-diffusion-v1/model_1_5.ckpt"
|
||||
},
|
||||
{
|
||||
"path": "configs/stable-diffusion/v1-inference.yaml"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "model_1_5.ckpt",
|
||||
"file_path": "models/ldm/stable-diffusion-v1",
|
||||
"file_url": "https://{username}:{password}@huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.ckpt",
|
||||
"hf_auth": true
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"stable_diffusion_1.4": {
|
||||
"name": "stable_diffusion",
|
||||
"type": "ckpt",
|
||||
"description": "Generalist AI image generating model. The baseline for all finetuned models.",
|
||||
"version": "1.4",
|
||||
"style": "generalist",
|
||||
"nsfw": false,
|
||||
"download_all": true,
|
||||
"requires": [
|
||||
"clip-vit-large-patch14"
|
||||
],
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/ldm/stable-diffusion-v1/model.ckpt",
|
||||
"md5sum": "c01059060130b8242849d86e97212c84"
|
||||
},
|
||||
{
|
||||
"path": "configs/stable-diffusion/v1-inference.yaml"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "model.ckpt",
|
||||
"file_path": "models/ldm/stable-diffusion-v1",
|
||||
"file_url": "https://www.googleapis.com/storage/v1/b/aai-blog-files/o/sd-v1-4.ckpt?alt=media"
|
||||
}
|
||||
],
|
||||
"alt_download": [
|
||||
{
|
||||
"file_name": "model.ckpt",
|
||||
"file_path": "models/ldm/stable-diffusion-v1",
|
||||
"file_url": "https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt",
|
||||
"hf_auth": true
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"waifu_diffusion": {
|
||||
"name": "waifu_diffusion",
|
||||
"type": "ckpt",
|
||||
"description": "Anime styled generations.",
|
||||
"version": "1.3",
|
||||
"style": "anime",
|
||||
"nsfw": false,
|
||||
"download_all": true,
|
||||
"requires": [
|
||||
"clip-vit-large-patch14"
|
||||
],
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/custom/waifu-diffusion.ckpt",
|
||||
"md5sum": "a2aa170e3f513b32a3fd8841656e0123"
|
||||
},
|
||||
{
|
||||
"path": "configs/stable-diffusion/v1-inference.yaml"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "waifu-diffusion.ckpt",
|
||||
"file_path": "models/custom",
|
||||
"file_url": "https://huggingface.co/hakurei/waifu-diffusion-v1-3/resolve/main/wd-v1-3-full.ckpt"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"Furry Epoch": {
|
||||
"name": "Furry Epoch",
|
||||
"type": "ckpt",
|
||||
"description": "Furry styled generations.",
|
||||
"version": "4",
|
||||
"style": "furry",
|
||||
"nsfw": false,
|
||||
"download_all": false,
|
||||
"requires": [
|
||||
"clip-vit-large-patch14"
|
||||
],
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/custom/furry-diffusion.ckpt",
|
||||
"md5sum": "f8ef45a295ef4966682f6e8fc2c6830d"
|
||||
},
|
||||
{
|
||||
"path": "configs/stable-diffusion/v1-inference.yaml"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "furry-diffusion.ckpt",
|
||||
"file_path": "models/custom",
|
||||
"file_url": "https://sexy.canine.wf/file/furry-ckpt/furry_epoch4.ckpt"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"Yiffy": {
|
||||
"name": "Yiffy",
|
||||
"type": "ckpt",
|
||||
"description": "Furry styled generations.",
|
||||
"version": "18",
|
||||
"style": "furry",
|
||||
"nsfw": false,
|
||||
"download_all": true,
|
||||
"requires": [
|
||||
"clip-vit-large-patch14"
|
||||
],
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/custom/yiffy.ckpt",
|
||||
"md5sum": "dbe25794e24af183565dc45e9ec99713"
|
||||
},
|
||||
{
|
||||
"path": "configs/stable-diffusion/v1-inference.yaml"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "yiffy.ckpt",
|
||||
"file_path": "models/custom",
|
||||
"file_url": "https://sexy.canine.wf/file/yiffy-ckpt/yiffy-e18.ckpt"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"Zack3D": {
|
||||
"name": "Zack3D",
|
||||
"type": "ckpt",
|
||||
"description": "Kink/NSFW oriented furry styled generations.",
|
||||
"version": "1",
|
||||
"style": "furry",
|
||||
"nsfw": true,
|
||||
"download_all": true,
|
||||
"requires": [
|
||||
"clip-vit-large-patch14"
|
||||
],
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/custom/Zack3D.ckpt",
|
||||
"md5sum": "aa944b1ecdaac60113027a0fdcda4f1b"
|
||||
},
|
||||
{
|
||||
"path": "configs/stable-diffusion/v1-inference.yaml"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "Zack3D.ckpt",
|
||||
"file_path": "models/custom",
|
||||
"file_url": "https://sexy.canine.wf/file/furry-ckpt/Zack3D_Kinky-v1.ckpt"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"trinart": {
|
||||
"name": "trinart",
|
||||
"type": "ckpt",
|
||||
"description": "Manga styled generations.",
|
||||
"version": "1",
|
||||
"style": "anime",
|
||||
"nsfw": false,
|
||||
"download_all": true,
|
||||
"requires": [
|
||||
"clip-vit-large-patch14"
|
||||
],
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/custom/trinart.ckpt"
|
||||
},
|
||||
{
|
||||
"path": "configs/stable-diffusion/v1-inference.yaml"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "trinart.ckpt",
|
||||
"file_path": "models/custom",
|
||||
"file_url": "https://huggingface.co/naclbit/trinart_stable_diffusion_v2/resolve/main/trinart2_step95000.ckpt"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"RealESRGAN_x4plus": {
|
||||
"name": "RealESRGAN_x4plus",
|
||||
"type": "realesrgan",
|
||||
"description": "Upscaler.",
|
||||
"version": "0.1.0",
|
||||
"style": "generalist",
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/realesrgan/RealESRGAN_x4plus.pth"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "RealESRGAN_x4plus.pth",
|
||||
"file_path": "models/realesrgan",
|
||||
"file_url": "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"RealESRGAN_x4plus_anime_6B": {
|
||||
"name": "RealESRGAN_x4plus_anime_6B",
|
||||
"type": "realesrgan",
|
||||
"description": "Anime focused upscaler.",
|
||||
"version": "0.2.2.4",
|
||||
"style": "anime",
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/realesrgan/RealESRGAN_x4plus_anime_6B.pth"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "RealESRGAN_x4plus_anime_6B.pth",
|
||||
"file_path": "models/realesrgan",
|
||||
"file_url": "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"GFPGAN": {
|
||||
"name": "GFPGAN",
|
||||
"type": "gfpgan",
|
||||
"description": "Face correction.",
|
||||
"version": "1.4",
|
||||
"style": "generalist",
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/gfpgan/GFPGANv1.4.pth"
|
||||
},
|
||||
{
|
||||
"path": "gfpgan/weights/detection_Resnet50_Final.pth"
|
||||
},
|
||||
{
|
||||
"path": "gfpgan/weights/parsing_parsenet.pth"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "GFPGANv1.4.pth",
|
||||
"file_path": "models/gfpgan",
|
||||
"file_url": "https://github.com/TencentARC/GFPGAN/releases/download/v1.3.4/GFPGANv1.4.pth"
|
||||
},
|
||||
{
|
||||
"file_name": "detection_Resnet50_Final.pth",
|
||||
"file_path": "./gfpgan/weights",
|
||||
"file_url": "https://github.com/xinntao/facexlib/releases/download/v0.1.0/detection_Resnet50_Final.pth"
|
||||
},
|
||||
{
|
||||
"file_name": "parsing_parsenet.pth",
|
||||
"file_path": "./gfpgan/weights",
|
||||
"file_url": "https://github.com/xinntao/facexlib/releases/download/v0.2.2/parsing_parsenet.pth"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"LDSR": {
|
||||
"name": "LDSR",
|
||||
"type": "ckpt",
|
||||
"description": "Upscaler.",
|
||||
"version": "1",
|
||||
"style": "generalist",
|
||||
"nsfw": false,
|
||||
"download_all": true,
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/ldsr/model.ckpt"
|
||||
},
|
||||
{
|
||||
"path": "models/ldsr/project.yaml"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "model.ckpt",
|
||||
"file_path": "models/ldsr",
|
||||
"file_url": "https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1"
|
||||
},
|
||||
{
|
||||
"file_name": "project.yaml",
|
||||
"file_path": "models/ldsr",
|
||||
"file_url": "https://heibox.uni-heidelberg.de/f/31a76b13ea27482981b4/?dl=1"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"BLIP": {
|
||||
"name": "BLIP",
|
||||
"type": "blip",
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/blip/model__base_caption.pth"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "model__base_caption.pth",
|
||||
"file_path": "models/blip",
|
||||
"file_url": "https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model*_base_caption.pth"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"ViT-L/14": {
|
||||
"name": "ViT-L/14",
|
||||
"type": "clip",
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/clip/ViT-L-14.pt"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "ViT-L-14.pt",
|
||||
"file_path": "./models/clip",
|
||||
"file_url": "https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"ViT-g-14": {
|
||||
"name": "ViT-g-14",
|
||||
"pretrained_name": "laion2b_s12b_b42k",
|
||||
"type": "open_clip",
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/clip/models--laion--CLIP-ViT-g-14-laion2B-s12B-b42K/"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "main",
|
||||
"file_path": "./models/clip/models--laion--CLIP-ViT-g-14-laion2B-s12B-b42K/refs",
|
||||
"file_content": "b36bdd32483debcf4ed2f918bdae1d4a46ee44b8"
|
||||
},
|
||||
{
|
||||
"file_name": "6aac683f899159946bc4ca15228bb7016f3cbb1a2c51f365cba0b23923f344da",
|
||||
"file_path": "./models/clip/models--laion--CLIP-ViT-g-14-laion2B-s12B-b42K/blobs",
|
||||
"file_url": "https://huggingface.co/laion/CLIP-ViT-g-14-laion2B-s12B-b42K/resolve/main/open_clip_pytorch_model.bin"
|
||||
},
|
||||
{
|
||||
"file_name": "open_clip_pytorch_model.bin",
|
||||
"file_path": "./models/clip/models--laion--CLIP-ViT-g-14-laion2B-s12B-b42K/snapshots/b36bdd32483debcf4ed2f918bdae1d4a46ee44b8",
|
||||
"symlink": "./models/clip/models--laion--CLIP-ViT-g-14-laion2B-s12B-b42K/blobs/6aac683f899159946bc4ca15228bb7016f3cbb1a2c51f365cba0b23923f344da"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"ViT-H-14": {
|
||||
"name": "ViT-H-14",
|
||||
"pretrained_name": "laion2b_s32b_b79k",
|
||||
"type": "open_clip",
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/clip/models--laion--CLIP-ViT-H-14-laion2B-s32B-b79K/"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "main",
|
||||
"file_path": "./models/clip/models--laion--CLIP-ViT-H-14-laion2B-s32B-b79K/refs",
|
||||
"file_content": "58a1e03a7acfacbe6b95ebc24ae0394eda6a14fc"
|
||||
},
|
||||
{
|
||||
"file_name": "9a78ef8e8c73fd0df621682e7a8e8eb36c6916cb3c16b291a082ecd52ab79cc4",
|
||||
"file_path": "./models/clip/models--laion--CLIP-ViT-H-14-laion2B-s32B-b79K/blobs",
|
||||
"file_url": "https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/resolve/main/open_clip_pytorch_model.bin"
|
||||
},
|
||||
{
|
||||
"file_name": "open_clip_pytorch_model.bin",
|
||||
"file_path": "./models/clip/models--laion--CLIP-ViT-H-14-laion2B-s32B-b79K/snapshots/58a1e03a7acfacbe6b95ebc24ae0394eda6a14fc",
|
||||
"symlink": "./models/clip/models--laion--CLIP-ViT-H-14-laion2B-s32B-b79K/blobs/9a78ef8e8c73fd0df621682e7a8e8eb36c6916cb3c16b291a082ecd52ab79cc4"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"diffusers_stable_diffusion": {
|
||||
"name": "diffusers_stable_diffusion",
|
||||
"type": "diffusers",
|
||||
"requires": [
|
||||
"clip-vit-large-patch14"
|
||||
],
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/diffusers/"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "diffusers_stable_diffusion",
|
||||
"file_url": "https://{username}:{password}@huggingface.co/CompVis/stable-diffusion-v1-4.git",
|
||||
"git": true,
|
||||
"hf_auth": true,
|
||||
"post_process": [
|
||||
{
|
||||
"delete": "models/diffusers/stable-diffusion-v1-4/.git"
|
||||
}
|
||||
]
|
||||
}
|
||||
]
|
||||
|
||||
},
|
||||
"available": false
|
||||
}
|
||||
}
|
||||
98
db_dep.json
Normal file
98
db_dep.json
Normal file
@ -0,0 +1,98 @@
|
||||
{
|
||||
"sd-concepts-library": {
|
||||
"type": "dependency",
|
||||
"optional": true,
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/custom/sd-concepts-library/"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "sd-concepts-library",
|
||||
"file_path": "./models/custom/sd-concepts-library/",
|
||||
"file_url": "https://github.com/sd-webui/sd-concepts-library/archive/refs/heads/main.zip",
|
||||
"unzip": true,
|
||||
"move_subfolder": "sd-concepts-library"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
},
|
||||
"clip-vit-large-patch14": {
|
||||
"type": "dependency",
|
||||
"optional": false,
|
||||
"config": {
|
||||
"files": [
|
||||
{
|
||||
"path": "models/clip-vit-large-patch14/config.json"
|
||||
},
|
||||
{
|
||||
"path": "models/clip-vit-large-patch14/merges.txt"
|
||||
},
|
||||
{
|
||||
"path": "models/clip-vit-large-patch14/preprocessor_config.json"
|
||||
},
|
||||
{
|
||||
"path": "models/clip-vit-large-patch14/pytorch_model.bin"
|
||||
},
|
||||
{
|
||||
"path": "models/clip-vit-large-patch14/special_tokens_map.json"
|
||||
},
|
||||
{
|
||||
"path": "models/clip-vit-large-patch14/tokenizer.json"
|
||||
},
|
||||
{
|
||||
"path": "models/clip-vit-large-patch14/tokenizer_config.json"
|
||||
},
|
||||
{
|
||||
"path": "models/clip-vit-large-patch14/vocab.json"
|
||||
}
|
||||
],
|
||||
"download": [
|
||||
{
|
||||
"file_name": "config.json",
|
||||
"file_path": "models/clip-vit-large-patch14",
|
||||
"file_url": "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/config.json"
|
||||
},
|
||||
{
|
||||
"file_name": "merges.txt",
|
||||
"file_path": "models/clip-vit-large-patch14",
|
||||
"file_url": "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/merges.txt"
|
||||
},
|
||||
{
|
||||
"file_name": "preprocessor_config.json",
|
||||
"file_path": "models/clip-vit-large-patch14",
|
||||
"file_url": "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/preprocessor_config.json"
|
||||
},
|
||||
{
|
||||
"file_name": "pytorch_model.bin",
|
||||
"file_path": "models/clip-vit-large-patch14",
|
||||
"file_url": "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/pytorch_model.bin"
|
||||
},
|
||||
{
|
||||
"file_name": "special_tokens_map.json",
|
||||
"file_path": "models/clip-vit-large-patch14",
|
||||
"file_url": "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/special_tokens_map.json"
|
||||
},
|
||||
{
|
||||
"file_name": "tokenizer.json",
|
||||
"file_path": "models/clip-vit-large-patch14",
|
||||
"file_url": "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/tokenizer.json"
|
||||
},
|
||||
{
|
||||
"file_name": "tokenizer_config.json",
|
||||
"file_path": "models/clip-vit-large-patch14",
|
||||
"file_url": "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/tokenizer_config.json"
|
||||
},
|
||||
{
|
||||
"file_name": "vocab.json",
|
||||
"file_path": "models/clip-vit-large-patch14",
|
||||
"file_url": "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/vocab.json"
|
||||
}
|
||||
]
|
||||
},
|
||||
"available": false
|
||||
}
|
||||
}
|
||||
92
docs/44.competition.md
Normal file
92
docs/44.competition.md
Normal file
@ -0,0 +1,92 @@
|
||||
# Textual inversion usage competition
|
||||
|
||||
We are hosting a competition where the community can showcase their most inventive use of textual inversion concepts in text-to-image or text-to-video.
|
||||
|
||||
Our compute cluster; `Nataili`, currently comprises of 3 nodes, two have 3090, the other has 2 x A5000.
|
||||
|
||||
We estimate `Nataili` can handle 12 concepts per hour, and we can add more workers if there is high demand.
|
||||
|
||||
Hopefully demand will be high, we want to train **hundreds** of new concepts!
|
||||
|
||||
# Schedule
|
||||
|
||||
2022/10/20 - Stage 1 begins, train concept command opened for usage
|
||||
|
||||
2022/10/22 12AM UTC - Stage 2 begins, text to image command opened for usage
|
||||
|
||||
2022/10/22 12PM UTC - Stage 1 ends, train concept command closed
|
||||
|
||||
2022/10/24 12PM UTC - Stage 2 ends, no more entries will be accepted
|
||||
|
||||
2022/10/24 6-12PM UTC - Winners announced
|
||||
|
||||
|
||||
# What does `most inventive use` mean?
|
||||
|
||||
Whatever you want it to mean! be creative! experiment!
|
||||
|
||||
There are several categories we will look at:
|
||||
|
||||
* anything that's particularly creative, ~ artistic ~ or a e s t h e t i c
|
||||
|
||||
|
||||
|
||||
* composition; meaning anything related to how big things are, their position, the angle, etc
|
||||
|
||||
* styling;
|
||||
|
||||
|
||||
|
||||
* `The Sims(TM): Stable Diffusion edition`
|
||||
|
||||
## So I can trai-
|
||||
|
||||
* Yes, as long as it's sfw
|
||||
|
||||
## `The Sims(TM): Stable Diffusion edition` ?
|
||||
|
||||
For this event the theme is “The Sims: Stable Diffusion edition”.
|
||||
|
||||
So we have selected a subset of [products from Amazon Berkely Objects dataset](https://github.com/sd-webui/abo).
|
||||
|
||||
Any other object is welcome too these are just a good source of data for this part of the competition.
|
||||
|
||||
Each product has images from multiple angles, the train concept command accepts up to 10 images, so choose the angles and modify backgrounds, experiment!
|
||||
|
||||
The goal with this category is to generate an image using the trained object, and the other categories apply, your imagination is the only limit! style a couch, try to make a BIG couch, try to make a couch on top of a mountain, try to make a vaporwave couch, anything!
|
||||
|
||||
# How do I train a concept using the discord bot?
|
||||
|
||||
Type `/trainconcept` then press tab to go through the fields
|
||||
|
||||
`Concept name` is just a name for your concept, it doesn't have to be a single word
|
||||
|
||||
`Placeholder` is what you will use in prompts to represent your concept
|
||||
Add `<` and `>` so it is unique, multiple words should be hyphenated
|
||||
|
||||
`Initializer` is used as the starting point for training your concept, so this should be a single word that represents your concept
|
||||
|
||||
Minimum 2 images. Squareish aspect ratios work best
|
||||
|
||||
|
||||
|
||||
|
||||
# How do I use the trained concept?
|
||||
|
||||
## Prompting with concepts
|
||||
|
||||
When your concept is trained you can use it in prompts.
|
||||
|
||||
`a cute <nvidiafu> as an astronaut`:
|
||||
|
||||
|
||||
|
||||
or `a green <green-couch> sitting on top of a floor, a 3D render, trending on polycount, minimalism, rendered in cinema4d`:
|
||||
|
||||
|
||||
|
||||
## Using concepts in the webui
|
||||
|
||||
The discord bot will give you a link to a `.zip` file, download this, extract it, and put the folder in `stable-diffusion-webui/models/custom/sd-concepts-library`
|
||||
|
||||
|
||||
@ -160,7 +160,7 @@ div.gallery:hover {
|
||||
/* Remove some empty spaces to make the UI more compact. */
|
||||
.css-18e3th9{
|
||||
padding-left: 10px;
|
||||
padding-right: 10px;
|
||||
padding-right: 30px;
|
||||
position: unset !important; /* Fixes the layout/page going up when an expander or another item is expanded and then collapsed */
|
||||
}
|
||||
.css-k1vhr4{
|
||||
@ -168,7 +168,7 @@ div.gallery:hover {
|
||||
}
|
||||
.css-ret2ud{
|
||||
padding-left: 10px;
|
||||
padding-right: 25px;
|
||||
padding-right: 30px;
|
||||
gap: initial;
|
||||
display: initial;
|
||||
}
|
||||
|
||||
@ -31,6 +31,8 @@ streamlit-option-menu==0.3.2
|
||||
streamlit_nested_layout==0.1.1
|
||||
streamlit-server-state==0.14.2
|
||||
streamlit-tensorboard==0.0.2
|
||||
streamlit-elements==0.1.* # used for the draggable dashboard and new UI design (WIP)
|
||||
streamlit-ace==0.1.1 # used to replace the text area on the prompt and also for the code editor tool.
|
||||
hydralit==1.0.14
|
||||
hydralit_components==1.0.10
|
||||
stqdm==0.0.4
|
||||
|
||||
@ -50,7 +50,7 @@ def layout():
|
||||
|
||||
with col1:
|
||||
st.title("General")
|
||||
st.session_state['defaults'].general.gpu = int(st.selectbox("GPU", device_list,
|
||||
st.session_state['defaults'].general.gpu = int(st.selectbox("GPU", device_list, index=st.session_state['defaults'].general.gpu,
|
||||
help=f"Select which GPU to use. Default: {device_list[0]}").split(":")[0])
|
||||
|
||||
st.session_state['defaults'].general.outdir = str(st.text_input("Output directory", value=st.session_state['defaults'].general.outdir,
|
||||
@ -208,15 +208,65 @@ def layout():
|
||||
with col4:
|
||||
st.title("Streamlit Config")
|
||||
|
||||
st.session_state["defaults"].general.streamlit_telemetry = st.checkbox("Enable Telemetry", value=st.session_state['defaults'].general.streamlit_telemetry,
|
||||
help="Enables or Disables streamlit telemetry. Default: False")
|
||||
st.session_state["streamlit_config"]["browser"]["gatherUsageStats"] = st.session_state["defaults"].general.streamlit_telemetry
|
||||
|
||||
default_theme_list = ["light", "dark"]
|
||||
st.session_state["defaults"].general.default_theme = st.selectbox("Default Theme", default_theme_list, index=default_theme_list.index(st.session_state['defaults'].general.default_theme),
|
||||
help="Defaut theme to use as base for streamlit. Default: dark")
|
||||
st.session_state["streamlit_config"]["theme"]["base"] = st.session_state["defaults"].general.default_theme
|
||||
|
||||
|
||||
if not st.session_state['defaults'].admin.hide_server_setting:
|
||||
with st.expander("Server", True):
|
||||
|
||||
st.session_state["streamlit_config"]['server']['headless'] = st.checkbox("Run Headless", help="If false, will attempt to open a browser window on start. \
|
||||
Default: false unless (1) we are on a Linux box where DISPLAY is unset, \
|
||||
or (2) we are running in the Streamlit Atom plugin.")
|
||||
|
||||
st.session_state["streamlit_config"]['server']['port'] = st.number_input("Port", value=st.session_state["streamlit_config"]['server']['port'],
|
||||
help="The port where the server will listen for browser connections. Default: 8501")
|
||||
|
||||
st.session_state["streamlit_config"]['server']['baseUrlPath'] = st.text_input("Base Url Path", value=st.session_state["streamlit_config"]['server']['baseUrlPath'],
|
||||
help="The base path for the URL where Streamlit should be served from. Default: '' ")
|
||||
|
||||
st.session_state["streamlit_config"]['server']['enableCORS'] = st.checkbox("Enable CORS", value=st.session_state['streamlit_config']['server']['enableCORS'],
|
||||
help="Enables support for Cross-Origin Request Sharing (CORS) protection, for added security. \
|
||||
Due to conflicts between CORS and XSRF, if `server.enableXsrfProtection` is on and `server.enableCORS` \
|
||||
is off at the same time, we will prioritize `server.enableXsrfProtection`. Default: true")
|
||||
|
||||
st.session_state["streamlit_config"]['server']['enableXsrfProtection'] = st.checkbox("Enable Xsrf Protection",
|
||||
value=st.session_state['streamlit_config']['server']['enableXsrfProtection'],
|
||||
help="Enables support for Cross-Site Request Forgery (XSRF) protection, \
|
||||
for added security. Due to conflicts between CORS and XSRF, \
|
||||
if `server.enableXsrfProtection` is on and `server.enableCORS` is off at \
|
||||
the same time, we will prioritize `server.enableXsrfProtection`. Default: true")
|
||||
|
||||
st.session_state["streamlit_config"]['server']['maxUploadSize'] = st.number_input("Max Upload Size", value=st.session_state["streamlit_config"]['server']['maxUploadSize'],
|
||||
help="Max size, in megabytes, for files uploaded with the file_uploader. Default: 200")
|
||||
|
||||
st.session_state["streamlit_config"]['server']['maxMessageSize'] = st.number_input("Max Message Size", value=st.session_state["streamlit_config"]['server']['maxUploadSize'],
|
||||
help="Max size, in megabytes, of messages that can be sent via the WebSocket connection. Default: 200")
|
||||
|
||||
st.session_state["streamlit_config"]['server']['enableWebsocketCompression'] = st.checkbox("Enable Websocket Compression",
|
||||
value=st.session_state["streamlit_config"]['server']['enableWebsocketCompression'],
|
||||
help=" Enables support for websocket compression. Default: false")
|
||||
if not st.session_state['defaults'].admin.hide_browser_setting:
|
||||
with st.expander("Browser", expanded=True):
|
||||
st.session_state["streamlit_config"]['browser']['serverAddress'] = st.text_input("Server Address",
|
||||
value=st.session_state["streamlit_config"]['browser']['serverAddress'] if "serverAddress" in st.session_state["streamlit_config"] else "localhost",
|
||||
help="Internet address where users should point their browsers in order \
|
||||
to connect to the app. Can be IP address or DNS name and path.\
|
||||
This is used to: - Set the correct URL for CORS and XSRF protection purposes. \
|
||||
- Show the URL on the terminal - Open the browser. Default: 'localhost'")
|
||||
|
||||
st.session_state["defaults"].general.streamlit_telemetry = st.checkbox("Enable Telemetry", value=st.session_state['defaults'].general.streamlit_telemetry,
|
||||
help="Enables or Disables streamlit telemetry. Default: False")
|
||||
st.session_state["streamlit_config"]["browser"]["gatherUsageStats"] = st.session_state["defaults"].general.streamlit_telemetry
|
||||
|
||||
st.session_state["streamlit_config"]['browser']['serverPort'] = st.number_input("Server Port", value=st.session_state["streamlit_config"]['browser']['serverPort'],
|
||||
help="Port where users should point their browsers in order to connect to the app. \
|
||||
This is used to: - Set the correct URL for CORS and XSRF protection purposes. \
|
||||
- Show the URL on the terminal - Open the browser \
|
||||
Default: whatever value is set in server.port.")
|
||||
|
||||
with col5:
|
||||
st.title("Huggingface")
|
||||
st.session_state["defaults"].general.huggingface_token = st.text_input("Huggingface Token", value=st.session_state['defaults'].general.huggingface_token, type="password",
|
||||
@ -225,6 +275,15 @@ def layout():
|
||||
and WILL NOT be share with us or anyone. You can get your access token \
|
||||
at https://huggingface.co/settings/tokens. Default: None")
|
||||
|
||||
st.title("Stable Horde")
|
||||
st.session_state["defaults"].general.stable_horde_api = st.text_input("Stable Horde Api", value=st.session_state["defaults"].general.stable_horde_api, type="password",
|
||||
help="First Register an account at https://stablehorde.net/register which will generate for you \
|
||||
an API key. Store that key somewhere safe. \n \
|
||||
If you do not want to register, you can use `0000000000` as api_key to connect anonymously.\
|
||||
However anonymous accounts have the lowest priority when there's too many concurrent requests! \
|
||||
To increase your priority you will need a unique API key and then to increase your Kudos \
|
||||
read more about them at https://dbzer0.com/blog/the-kudos-based-economy-for-the-koboldai-horde/.")
|
||||
|
||||
with txt2img_tab:
|
||||
col1, col2, col3, col4, col5 = st.columns(5, gap='medium')
|
||||
|
||||
|
||||
97
scripts/barfi_baklavajs.py
Normal file
97
scripts/barfi_baklavajs.py
Normal file
@ -0,0 +1,97 @@
|
||||
# This file is part of sygil-webui (https://github.com/Sygil-Dev/sandbox-webui/).
|
||||
|
||||
# Copyright 2022 Sygil-Dev team.
|
||||
# This program is free software: you can redistribute it and/or modify
|
||||
# it under the terms of the GNU Affero General Public License as published by
|
||||
# the Free Software Foundation, either version 3 of the License, or
|
||||
# (at your option) any later version.
|
||||
|
||||
# This program is distributed in the hope that it will be useful,
|
||||
# but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
# GNU Affero General Public License for more details.
|
||||
|
||||
# You should have received a copy of the GNU Affero General Public License
|
||||
# along with this program. If not, see <http://www.gnu.org/licenses/>.
|
||||
# base webui import and utils.
|
||||
#from sd_utils import *
|
||||
from sd_utils import *
|
||||
# streamlit imports
|
||||
|
||||
#streamlit components section
|
||||
|
||||
#other imports
|
||||
import os, time, requests
|
||||
import sys
|
||||
|
||||
from barfi import st_barfi, barfi_schemas, Block
|
||||
|
||||
# Temp imports
|
||||
|
||||
# end of imports
|
||||
#---------------------------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def layout():
|
||||
#st.info("Under Construction. :construction_worker:")
|
||||
|
||||
#from barfi import st_barfi, Block
|
||||
|
||||
#add = Block(name='Addition')
|
||||
#sub = Block(name='Subtraction')
|
||||
#mul = Block(name='Multiplication')
|
||||
#div = Block(name='Division')
|
||||
|
||||
#barfi_result = st_barfi(base_blocks= [add, sub, mul, div])
|
||||
# or if you want to use a category to organise them in the frontend sub-menu
|
||||
#barfi_result = st_barfi(base_blocks= {'Op 1': [add, sub], 'Op 2': [mul, div]})
|
||||
|
||||
col1, col2, col3 = st.columns([1, 8, 1])
|
||||
|
||||
from barfi import st_barfi, barfi_schemas, Block
|
||||
|
||||
|
||||
with col2:
|
||||
feed = Block(name='Feed')
|
||||
feed.add_output()
|
||||
def feed_func(self):
|
||||
self.set_interface(name='Output 1', value=4)
|
||||
feed.add_compute(feed_func)
|
||||
|
||||
splitter = Block(name='Splitter')
|
||||
splitter.add_input()
|
||||
splitter.add_output()
|
||||
splitter.add_output()
|
||||
def splitter_func(self):
|
||||
in_1 = self.get_interface(name='Input 1')
|
||||
value = (in_1/2)
|
||||
self.set_interface(name='Output 1', value=value)
|
||||
self.set_interface(name='Output 2', value=value)
|
||||
splitter.add_compute(splitter_func)
|
||||
|
||||
mixer = Block(name='Mixer')
|
||||
mixer.add_input()
|
||||
mixer.add_input()
|
||||
mixer.add_output()
|
||||
def mixer_func(self):
|
||||
in_1 = self.get_interface(name='Input 1')
|
||||
in_2 = self.get_interface(name='Input 2')
|
||||
value = (in_1 + in_2)
|
||||
self.set_interface(name='Output 1', value=value)
|
||||
mixer.add_compute(mixer_func)
|
||||
|
||||
result = Block(name='Result')
|
||||
result.add_input()
|
||||
def result_func(self):
|
||||
in_1 = self.get_interface(name='Input 1')
|
||||
result.add_compute(result_func)
|
||||
|
||||
load_schema = st.selectbox('Select a saved schema:', barfi_schemas())
|
||||
|
||||
compute_engine = st.checkbox('Activate barfi compute engine', value=False)
|
||||
|
||||
barfi_result = st_barfi(base_blocks=[feed, result, mixer, splitter],
|
||||
compute_engine=compute_engine, load_schema=load_schema)
|
||||
|
||||
if barfi_result:
|
||||
st.write(barfi_result)
|
||||
@ -33,6 +33,7 @@ from ldm.models.diffusion.plms import PLMSSampler
|
||||
|
||||
# streamlit components
|
||||
from custom_components import sygil_suggestions
|
||||
from streamlit_drawable_canvas import st_canvas
|
||||
|
||||
# Temp imports
|
||||
|
||||
@ -381,7 +382,7 @@ def layout():
|
||||
|
||||
|
||||
# creating the page layout using columns
|
||||
col1_img2img_layout, col2_img2img_layout, col3_img2img_layout = st.columns([1,2,2], gap="medium")
|
||||
col1_img2img_layout, col2_img2img_layout, col3_img2img_layout = st.columns([2,4,4], gap="medium")
|
||||
|
||||
with col1_img2img_layout:
|
||||
# If we have custom models available on the "models/custom"
|
||||
@ -426,7 +427,7 @@ def layout():
|
||||
mask_expander = st.empty()
|
||||
with mask_expander.expander("Mask"):
|
||||
mask_mode_list = ["Mask", "Inverted mask", "Image alpha"]
|
||||
mask_mode = st.selectbox("Mask Mode", mask_mode_list,
|
||||
mask_mode = st.selectbox("Mask Mode", mask_mode_list, index=st.session_state["defaults"].img2img.mask_mode,
|
||||
help="Select how you want your image to be masked.\"Mask\" modifies the image where the mask is white.\n\
|
||||
\"Inverted mask\" modifies the image where the mask is black. \"Image alpha\" modifies the image where the image is transparent."
|
||||
)
|
||||
@ -434,15 +435,32 @@ def layout():
|
||||
|
||||
|
||||
noise_mode_list = ["Seed", "Find Noise", "Matched Noise", "Find+Matched Noise"]
|
||||
noise_mode = st.selectbox(
|
||||
"Noise Mode", noise_mode_list,
|
||||
help=""
|
||||
)
|
||||
noise_mode = noise_mode_list.index(noise_mode)
|
||||
noise_mode = st.selectbox("Noise Mode", noise_mode_list, index=noise_mode_list.index(st.session_state['defaults'].img2img.noise_mode), help="")
|
||||
#noise_mode = noise_mode_list.index(noise_mode)
|
||||
find_noise_steps = st.number_input("Find Noise Steps", value=st.session_state['defaults'].img2img.find_noise_steps.value,
|
||||
min_value=st.session_state['defaults'].img2img.find_noise_steps.min_value,
|
||||
step=st.session_state['defaults'].img2img.find_noise_steps.step)
|
||||
|
||||
# Specify canvas parameters in application
|
||||
drawing_mode = st.selectbox(
|
||||
"Drawing tool:",
|
||||
(
|
||||
"freedraw",
|
||||
"transform",
|
||||
#"line",
|
||||
"rect",
|
||||
"circle",
|
||||
#"polygon",
|
||||
),
|
||||
)
|
||||
|
||||
stroke_width = st.slider("Stroke width: ", 1, 100, 50)
|
||||
stroke_color = st.color_picker("Stroke color hex: ", value="#EEEEEE")
|
||||
bg_color = st.color_picker("Background color hex: ", "#7B6E6E")
|
||||
|
||||
display_toolbar = st.checkbox("Display toolbar", True)
|
||||
#realtime_update = st.checkbox("Update in realtime", True)
|
||||
|
||||
with st.expander("Batch Options"):
|
||||
st.session_state["batch_count"] = st.number_input("Batch count.", value=st.session_state['defaults'].img2img.batch_count.value,
|
||||
help="How many iterations or batches of images to generate in total.")
|
||||
@ -527,7 +545,9 @@ def layout():
|
||||
upscaling_method_list.append("LDSR")
|
||||
|
||||
st.session_state["upscaling_method"] = st.selectbox("Upscaling Method", upscaling_method_list,
|
||||
index=upscaling_method_list.index(st.session_state['defaults'].general.upscaling_method))
|
||||
index=upscaling_method_list.index(st.session_state['defaults'].general.upscaling_method)
|
||||
if st.session_state['defaults'].general.upscaling_method in upscaling_method_list
|
||||
else 0)
|
||||
|
||||
if st.session_state["RealESRGAN_available"]:
|
||||
with st.expander("RealESRGAN"):
|
||||
@ -583,55 +603,63 @@ def layout():
|
||||
editor_image = st.empty()
|
||||
st.session_state["editor_image"] = editor_image
|
||||
|
||||
st.form_submit_button("Refresh")
|
||||
|
||||
#if "canvas" not in st.session_state:
|
||||
st.session_state["canvas"] = st.empty()
|
||||
|
||||
masked_image_holder = st.empty()
|
||||
image_holder = st.empty()
|
||||
|
||||
st.form_submit_button("Refresh")
|
||||
|
||||
uploaded_images = st.file_uploader(
|
||||
"Upload Image", accept_multiple_files=False, type=["png", "jpg", "jpeg", "webp", 'jfif'],
|
||||
help="Upload an image which will be used for the image to image generation.",
|
||||
)
|
||||
if uploaded_images:
|
||||
image = Image.open(uploaded_images).convert('RGBA')
|
||||
image = Image.open(uploaded_images).convert('RGB')
|
||||
new_img = image.resize((width, height))
|
||||
image_holder.image(new_img)
|
||||
#image_holder.image(new_img)
|
||||
|
||||
mask_holder = st.empty()
|
||||
#mask_holder = st.empty()
|
||||
|
||||
uploaded_masks = st.file_uploader(
|
||||
"Upload Mask", accept_multiple_files=False, type=["png", "jpg", "jpeg", "webp", 'jfif'],
|
||||
help="Upload an mask image which will be used for masking the image to image generation.",
|
||||
#uploaded_masks = st.file_uploader(
|
||||
#"Upload Mask", accept_multiple_files=False, type=["png", "jpg", "jpeg", "webp", 'jfif'],
|
||||
#help="Upload an mask image which will be used for masking the image to image generation.",
|
||||
#)
|
||||
|
||||
#
|
||||
# Create a canvas component
|
||||
with st.session_state["canvas"]:
|
||||
st.session_state["uploaded_masks"] = st_canvas(
|
||||
fill_color="rgba(255, 165, 0, 0.3)", # Fixed fill color with some opacity
|
||||
stroke_width=stroke_width,
|
||||
stroke_color=stroke_color,
|
||||
background_color=bg_color,
|
||||
background_image=image if uploaded_images else None,
|
||||
update_streamlit=True,
|
||||
width=width,
|
||||
height=height,
|
||||
drawing_mode=drawing_mode,
|
||||
initial_drawing=st.session_state["uploaded_masks"].json_data if "uploaded_masks" in st.session_state else None,
|
||||
display_toolbar= display_toolbar,
|
||||
key="full_app",
|
||||
)
|
||||
if uploaded_masks:
|
||||
mask_expander.expander("Mask", expanded=True)
|
||||
mask = Image.open(uploaded_masks)
|
||||
if mask.mode == "RGBA":
|
||||
mask = mask.convert('RGBA')
|
||||
background = Image.new('RGBA', mask.size, (0, 0, 0))
|
||||
mask = Image.alpha_composite(background, mask)
|
||||
mask = mask.resize((width, height))
|
||||
mask_holder.image(mask)
|
||||
|
||||
if uploaded_images and uploaded_masks:
|
||||
if mask_mode != 2:
|
||||
final_img = new_img.copy()
|
||||
alpha_layer = mask.convert('L')
|
||||
strength = st.session_state["denoising_strength"]
|
||||
if mask_mode == 0:
|
||||
alpha_layer = ImageOps.invert(alpha_layer)
|
||||
alpha_layer = alpha_layer.point(lambda a: a * strength)
|
||||
alpha_layer = ImageOps.invert(alpha_layer)
|
||||
elif mask_mode == 1:
|
||||
alpha_layer = alpha_layer.point(lambda a: a * strength)
|
||||
alpha_layer = ImageOps.invert(alpha_layer)
|
||||
#try:
|
||||
##print (type(st.session_state["uploaded_masks"]))
|
||||
#if st.session_state["uploaded_masks"] != None:
|
||||
#mask_expander.expander("Mask", expanded=True)
|
||||
#mask = Image.fromarray(st.session_state["uploaded_masks"].image_data)
|
||||
|
||||
final_img.putalpha(alpha_layer)
|
||||
|
||||
with masked_image_holder.container():
|
||||
st.text("Masked Image Preview")
|
||||
st.image(final_img)
|
||||
#st.image(mask)
|
||||
|
||||
#if mask.mode == "RGBA":
|
||||
#mask = mask.convert('RGBA')
|
||||
#background = Image.new('RGBA', mask.size, (0, 0, 0))
|
||||
#mask = Image.alpha_composite(background, mask)
|
||||
#mask = mask.resize((width, height))
|
||||
#except AttributeError:
|
||||
#pass
|
||||
|
||||
with col3_img2img_layout:
|
||||
result_tab = st.tabs(["Result"])
|
||||
@ -645,7 +673,6 @@ def layout():
|
||||
st.session_state["progress_bar_text"] = st.empty()
|
||||
st.session_state["progress_bar"] = st.empty()
|
||||
|
||||
|
||||
message = st.empty()
|
||||
|
||||
#if uploaded_images:
|
||||
@ -666,14 +693,17 @@ def layout():
|
||||
CustomModel_available=server_state["CustomModel_available"], custom_model=st.session_state["custom_model"])
|
||||
|
||||
if uploaded_images:
|
||||
image = Image.open(uploaded_images).convert('RGBA')
|
||||
new_img = image.resize((width, height))
|
||||
#img_array = np.array(image) # if you want to pass it to OpenCV
|
||||
#image = Image.fromarray(image).convert('RGBA')
|
||||
#new_img = image.resize((width, height))
|
||||
###img_array = np.array(image) # if you want to pass it to OpenCV
|
||||
#image_holder.image(new_img)
|
||||
new_mask = None
|
||||
if uploaded_masks:
|
||||
mask = Image.open(uploaded_masks).convert('RGBA')
|
||||
|
||||
if st.session_state["uploaded_masks"]:
|
||||
mask = Image.fromarray(st.session_state["uploaded_masks"].image_data)
|
||||
new_mask = mask.resize((width, height))
|
||||
|
||||
#masked_image_holder.image(new_mask)
|
||||
try:
|
||||
output_images, seed, info, stats = img2img(prompt=prompt, init_info=new_img, init_info_mask=new_mask, mask_mode=mask_mode,
|
||||
mask_restore=img2img_mask_restore, ddim_steps=st.session_state["sampling_steps"],
|
||||
|
||||
@ -367,10 +367,7 @@ def layout():
|
||||
col1, col2 = st.columns([1, 4], gap="large")
|
||||
|
||||
with col1:
|
||||
#url = st.text_area("Input Text","")
|
||||
#url = st.text_input("Input Text","", placeholder="A corgi wearing a top hat as an oil painting.")
|
||||
#st.subheader("Input Image")
|
||||
st.session_state["uploaded_image"] = st.file_uploader('Input Image', type=['png', 'jpg', 'jpeg', 'jfif'], accept_multiple_files=True)
|
||||
st.session_state["uploaded_image"] = st.file_uploader('Input Image', type=['png', 'jpg', 'jpeg', 'jfif', 'webp'], accept_multiple_files=True)
|
||||
|
||||
with st.expander("CLIP models", expanded=True):
|
||||
st.session_state["ViT-L/14"] = st.checkbox("ViT-L/14", value=True, help="ViT-L/14 model.")
|
||||
|
||||
@ -1,551 +0,0 @@
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
import k_diffusion as K
|
||||
import tqdm
|
||||
from contextlib import contextmanager, nullcontext
|
||||
import skimage
|
||||
import numpy as np
|
||||
import PIL
|
||||
import torch
|
||||
from einops import rearrange
|
||||
from ldm.models.diffusion.ddim import DDIMSampler
|
||||
from ldm.models.diffusion.kdiffusion import CFGMaskedDenoiser, KDiffusionSampler
|
||||
from ldm.models.diffusion.plms import PLMSSampler
|
||||
from nataili.util.cache import torch_gc
|
||||
from nataili.util.check_prompt_length import check_prompt_length
|
||||
from nataili.util.get_next_sequence_number import get_next_sequence_number
|
||||
from nataili.util.image_grid import image_grid
|
||||
from nataili.util.load_learned_embed_in_clip import load_learned_embed_in_clip
|
||||
from nataili.util.save_sample import save_sample
|
||||
from nataili.util.seed_to_int import seed_to_int
|
||||
from slugify import slugify
|
||||
import PIL
|
||||
|
||||
|
||||
class img2img:
|
||||
def __init__(self, model, device, output_dir, save_extension='jpg',
|
||||
output_file_path=False, load_concepts=False, concepts_dir=None,
|
||||
verify_input=True, auto_cast=True):
|
||||
self.model = model
|
||||
self.output_dir = output_dir
|
||||
self.output_file_path = output_file_path
|
||||
self.save_extension = save_extension
|
||||
self.load_concepts = load_concepts
|
||||
self.concepts_dir = concepts_dir
|
||||
self.verify_input = verify_input
|
||||
self.auto_cast = auto_cast
|
||||
self.device = device
|
||||
self.comments = []
|
||||
self.output_images = []
|
||||
self.info = ''
|
||||
self.stats = ''
|
||||
self.images = []
|
||||
|
||||
def create_random_tensors(self, shape, seeds):
|
||||
xs = []
|
||||
for seed in seeds:
|
||||
torch.manual_seed(seed)
|
||||
|
||||
# randn results depend on device; gpu and cpu get different results for same seed;
|
||||
# the way I see it, it's better to do this on CPU, so that everyone gets same result;
|
||||
# but the original script had it like this so i do not dare change it for now because
|
||||
# it will break everyone's seeds.
|
||||
xs.append(torch.randn(shape, device=self.device))
|
||||
x = torch.stack(xs)
|
||||
return x
|
||||
|
||||
def process_prompt_tokens(self, prompt_tokens):
|
||||
# compviz codebase
|
||||
tokenizer = self.model.cond_stage_model.tokenizer
|
||||
text_encoder = self.model.cond_stage_model.transformer
|
||||
|
||||
# diffusers codebase
|
||||
#tokenizer = pipe.tokenizer
|
||||
#text_encoder = pipe.text_encoder
|
||||
|
||||
ext = ('.pt', '.bin')
|
||||
for token_name in prompt_tokens:
|
||||
embedding_path = os.path.join(self.concepts_dir, token_name)
|
||||
if os.path.exists(embedding_path):
|
||||
for files in os.listdir(embedding_path):
|
||||
if files.endswith(ext):
|
||||
load_learned_embed_in_clip(f"{os.path.join(embedding_path, files)}", text_encoder, tokenizer, f"<{token_name}>")
|
||||
else:
|
||||
print(f"Concept {token_name} not found in {self.concepts_dir}")
|
||||
del tokenizer, text_encoder
|
||||
return
|
||||
del tokenizer, text_encoder
|
||||
|
||||
def resize_image(self, resize_mode, im, width, height):
|
||||
LANCZOS = (PIL.Image.Resampling.LANCZOS if hasattr(PIL.Image, 'Resampling') else PIL.Image.LANCZOS)
|
||||
if resize_mode == "resize":
|
||||
res = im.resize((width, height), resample=LANCZOS)
|
||||
elif resize_mode == "crop":
|
||||
ratio = width / height
|
||||
src_ratio = im.width / im.height
|
||||
|
||||
src_w = width if ratio > src_ratio else im.width * height // im.height
|
||||
src_h = height if ratio <= src_ratio else im.height * width // im.width
|
||||
|
||||
resized = im.resize((src_w, src_h), resample=LANCZOS)
|
||||
res = PIL.Image.new("RGBA", (width, height))
|
||||
res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
|
||||
else:
|
||||
ratio = width / height
|
||||
src_ratio = im.width / im.height
|
||||
|
||||
src_w = width if ratio < src_ratio else im.width * height // im.height
|
||||
src_h = height if ratio >= src_ratio else im.height * width // im.width
|
||||
|
||||
resized = im.resize((src_w, src_h), resample=LANCZOS)
|
||||
res = PIL.Image.new("RGBA", (width, height))
|
||||
res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
|
||||
|
||||
if ratio < src_ratio:
|
||||
fill_height = height // 2 - src_h // 2
|
||||
res.paste(resized.resize((width, fill_height), box=(0, 0, width, 0)), box=(0, 0))
|
||||
res.paste(resized.resize((width, fill_height), box=(0, resized.height, width, resized.height)), box=(0, fill_height + src_h))
|
||||
elif ratio > src_ratio:
|
||||
fill_width = width // 2 - src_w // 2
|
||||
res.paste(resized.resize((fill_width, height), box=(0, 0, 0, height)), box=(0, 0))
|
||||
res.paste(resized.resize((fill_width, height), box=(resized.width, 0, resized.width, height)), box=(fill_width + src_w, 0))
|
||||
|
||||
return res
|
||||
|
||||
#
|
||||
# helper fft routines that keep ortho normalization and auto-shift before and after fft
|
||||
def _fft2(self, data):
|
||||
if data.ndim > 2: # has channels
|
||||
out_fft = np.zeros((data.shape[0], data.shape[1], data.shape[2]), dtype=np.complex128)
|
||||
for c in range(data.shape[2]):
|
||||
c_data = data[:,:,c]
|
||||
out_fft[:,:,c] = np.fft.fft2(np.fft.fftshift(c_data),norm="ortho")
|
||||
out_fft[:,:,c] = np.fft.ifftshift(out_fft[:,:,c])
|
||||
else: # one channel
|
||||
out_fft = np.zeros((data.shape[0], data.shape[1]), dtype=np.complex128)
|
||||
out_fft[:,:] = np.fft.fft2(np.fft.fftshift(data),norm="ortho")
|
||||
out_fft[:,:] = np.fft.ifftshift(out_fft[:,:])
|
||||
|
||||
return out_fft
|
||||
|
||||
def _ifft2(self, data):
|
||||
if data.ndim > 2: # has channels
|
||||
out_ifft = np.zeros((data.shape[0], data.shape[1], data.shape[2]), dtype=np.complex128)
|
||||
for c in range(data.shape[2]):
|
||||
c_data = data[:,:,c]
|
||||
out_ifft[:,:,c] = np.fft.ifft2(np.fft.fftshift(c_data),norm="ortho")
|
||||
out_ifft[:,:,c] = np.fft.ifftshift(out_ifft[:,:,c])
|
||||
else: # one channel
|
||||
out_ifft = np.zeros((data.shape[0], data.shape[1]), dtype=np.complex128)
|
||||
out_ifft[:,:] = np.fft.ifft2(np.fft.fftshift(data),norm="ortho")
|
||||
out_ifft[:,:] = np.fft.ifftshift(out_ifft[:,:])
|
||||
|
||||
return out_ifft
|
||||
|
||||
def _get_gaussian_window(self, width, height, std=3.14, mode=0):
|
||||
|
||||
window_scale_x = float(width / min(width, height))
|
||||
window_scale_y = float(height / min(width, height))
|
||||
|
||||
window = np.zeros((width, height))
|
||||
x = (np.arange(width) / width * 2. - 1.) * window_scale_x
|
||||
for y in range(height):
|
||||
fy = (y / height * 2. - 1.) * window_scale_y
|
||||
if mode == 0:
|
||||
window[:, y] = np.exp(-(x**2+fy**2) * std)
|
||||
else:
|
||||
window[:, y] = (1/((x**2+1.) * (fy**2+1.))) ** (std/3.14) # hey wait a minute that's not gaussian
|
||||
|
||||
return window
|
||||
|
||||
def _get_masked_window_rgb(self, np_mask_grey, hardness=1.):
|
||||
np_mask_rgb = np.zeros((np_mask_grey.shape[0], np_mask_grey.shape[1], 3))
|
||||
if hardness != 1.:
|
||||
hardened = np_mask_grey[:] ** hardness
|
||||
else:
|
||||
hardened = np_mask_grey[:]
|
||||
for c in range(3):
|
||||
np_mask_rgb[:,:,c] = hardened[:]
|
||||
return np_mask_rgb
|
||||
|
||||
def get_matched_noise(self, _np_src_image, np_mask_rgb, noise_q, color_variation):
|
||||
"""
|
||||
Explanation:
|
||||
Getting good results in/out-painting with stable diffusion can be challenging.
|
||||
Although there are simpler effective solutions for in-painting, out-painting can be especially challenging because there is no color data
|
||||
in the masked area to help prompt the generator. Ideally, even for in-painting we'd like work effectively without that data as well.
|
||||
Provided here is my take on a potential solution to this problem.
|
||||
|
||||
By taking a fourier transform of the masked src img we get a function that tells us the presence and orientation of each feature scale in the unmasked src.
|
||||
Shaping the init/seed noise for in/outpainting to the same distribution of feature scales, orientations, and positions increases output coherence
|
||||
by helping keep features aligned. This technique is applicable to any continuous generation task such as audio or video, each of which can
|
||||
be conceptualized as a series of out-painting steps where the last half of the input "frame" is erased. For multi-channel data such as color
|
||||
or stereo sound the "color tone" or histogram of the seed noise can be matched to improve quality (using scikit-image currently)
|
||||
This method is quite robust and has the added benefit of being fast independently of the size of the out-painted area.
|
||||
The effects of this method include things like helping the generator integrate the pre-existing view distance and camera angle.
|
||||
|
||||
Carefully managing color and brightness with histogram matching is also essential to achieving good coherence.
|
||||
|
||||
noise_q controls the exponent in the fall-off of the distribution can be any positive number, lower values means higher detail (range > 0, default 1.)
|
||||
color_variation controls how much freedom is allowed for the colors/palette of the out-painted area (range 0..1, default 0.01)
|
||||
This code is provided as is under the Unlicense (https://unlicense.org/)
|
||||
Although you have no obligation to do so, if you found this code helpful please find it in your heart to credit me [parlance-zz].
|
||||
|
||||
Questions or comments can be sent to parlance@fifth-harmonic.com (https://github.com/parlance-zz/)
|
||||
This code is part of a new branch of a discord bot I am working on integrating with diffusers (https://github.com/parlance-zz/g-diffuser-bot)
|
||||
|
||||
"""
|
||||
|
||||
global DEBUG_MODE
|
||||
global TMP_ROOT_PATH
|
||||
|
||||
width = _np_src_image.shape[0]
|
||||
height = _np_src_image.shape[1]
|
||||
num_channels = _np_src_image.shape[2]
|
||||
|
||||
np_src_image = _np_src_image[:] * (1. - np_mask_rgb)
|
||||
np_mask_grey = (np.sum(np_mask_rgb, axis=2)/3.)
|
||||
np_src_grey = (np.sum(np_src_image, axis=2)/3.)
|
||||
all_mask = np.ones((width, height), dtype=bool)
|
||||
img_mask = np_mask_grey > 1e-6
|
||||
ref_mask = np_mask_grey < 1e-3
|
||||
|
||||
windowed_image = _np_src_image * (1.-self._get_masked_window_rgb(np_mask_grey))
|
||||
windowed_image /= np.max(windowed_image)
|
||||
windowed_image += np.average(_np_src_image) * np_mask_rgb# / (1.-np.average(np_mask_rgb)) # rather than leave the masked area black, we get better results from fft by filling the average unmasked color
|
||||
#windowed_image += np.average(_np_src_image) * (np_mask_rgb * (1.- np_mask_rgb)) / (1.-np.average(np_mask_rgb)) # compensate for darkening across the mask transition area
|
||||
#_save_debug_img(windowed_image, "windowed_src_img")
|
||||
|
||||
src_fft = self._fft2(windowed_image) # get feature statistics from masked src img
|
||||
src_dist = np.absolute(src_fft)
|
||||
src_phase = src_fft / src_dist
|
||||
#_save_debug_img(src_dist, "windowed_src_dist")
|
||||
|
||||
noise_window = self._get_gaussian_window(width, height, mode=1) # start with simple gaussian noise
|
||||
noise_rgb = np.random.random_sample((width, height, num_channels))
|
||||
noise_grey = (np.sum(noise_rgb, axis=2)/3.)
|
||||
noise_rgb *= color_variation # the colorfulness of the starting noise is blended to greyscale with a parameter
|
||||
for c in range(num_channels):
|
||||
noise_rgb[:,:,c] += (1. - color_variation) * noise_grey
|
||||
|
||||
noise_fft = self._fft2(noise_rgb)
|
||||
for c in range(num_channels):
|
||||
noise_fft[:,:,c] *= noise_window
|
||||
noise_rgb = np.real(self._ifft2(noise_fft))
|
||||
shaped_noise_fft = self._fft2(noise_rgb)
|
||||
shaped_noise_fft[:,:,:] = np.absolute(shaped_noise_fft[:,:,:])**2 * (src_dist ** noise_q) * src_phase # perform the actual shaping
|
||||
|
||||
brightness_variation = 0.#color_variation # todo: temporarily tieing brightness variation to color variation for now
|
||||
contrast_adjusted_np_src = _np_src_image[:] * (brightness_variation + 1.) - brightness_variation * 2.
|
||||
|
||||
# scikit-image is used for histogram matching, very convenient!
|
||||
shaped_noise = np.real(self._ifft2(shaped_noise_fft))
|
||||
shaped_noise -= np.min(shaped_noise)
|
||||
shaped_noise /= np.max(shaped_noise)
|
||||
shaped_noise[img_mask,:] = skimage.exposure.match_histograms(shaped_noise[img_mask,:]**1., contrast_adjusted_np_src[ref_mask,:], channel_axis=1)
|
||||
shaped_noise = _np_src_image[:] * (1. - np_mask_rgb) + shaped_noise * np_mask_rgb
|
||||
#_save_debug_img(shaped_noise, "shaped_noise")
|
||||
|
||||
matched_noise = np.zeros((width, height, num_channels))
|
||||
matched_noise = shaped_noise[:]
|
||||
#matched_noise[all_mask,:] = skimage.exposure.match_histograms(shaped_noise[all_mask,:], _np_src_image[ref_mask,:], channel_axis=1)
|
||||
#matched_noise = _np_src_image[:] * (1. - np_mask_rgb) + matched_noise * np_mask_rgb
|
||||
|
||||
#_save_debug_img(matched_noise, "matched_noise")
|
||||
|
||||
"""
|
||||
todo:
|
||||
color_variation doesnt have to be a single number, the overall color tone of the out-painted area could be param controlled
|
||||
"""
|
||||
|
||||
return np.clip(matched_noise, 0., 1.)
|
||||
|
||||
def find_noise_for_image(self, model, device, init_image, prompt, steps=200, cond_scale=2.0, verbose=False, normalize=False, generation_callback=None):
|
||||
image = np.array(init_image).astype(np.float32) / 255.0
|
||||
image = image[None].transpose(0, 3, 1, 2)
|
||||
image = torch.from_numpy(image)
|
||||
image = 2. * image - 1.
|
||||
image = image.to(device)
|
||||
x = model.get_first_stage_encoding(model.encode_first_stage(image))
|
||||
|
||||
uncond = model.get_learned_conditioning([''])
|
||||
cond = model.get_learned_conditioning([prompt])
|
||||
|
||||
s_in = x.new_ones([x.shape[0]])
|
||||
dnw = K.external.CompVisDenoiser(model)
|
||||
sigmas = dnw.get_sigmas(steps).flip(0)
|
||||
|
||||
if verbose:
|
||||
print(sigmas)
|
||||
|
||||
for i in tqdm.trange(1, len(sigmas)):
|
||||
x_in = torch.cat([x] * 2)
|
||||
sigma_in = torch.cat([sigmas[i - 1] * s_in] * 2)
|
||||
cond_in = torch.cat([uncond, cond])
|
||||
|
||||
c_out, c_in = [K.utils.append_dims(k, x_in.ndim) for k in dnw.get_scalings(sigma_in)]
|
||||
|
||||
if i == 1:
|
||||
t = dnw.sigma_to_t(torch.cat([sigmas[i] * s_in] * 2))
|
||||
else:
|
||||
t = dnw.sigma_to_t(sigma_in)
|
||||
|
||||
eps = model.apply_model(x_in * c_in, t, cond=cond_in)
|
||||
denoised_uncond, denoised_cond = (x_in + eps * c_out).chunk(2)
|
||||
|
||||
denoised = denoised_uncond + (denoised_cond - denoised_uncond) * cond_scale
|
||||
|
||||
if i == 1:
|
||||
d = (x - denoised) / (2 * sigmas[i])
|
||||
else:
|
||||
d = (x - denoised) / sigmas[i - 1]
|
||||
|
||||
dt = sigmas[i] - sigmas[i - 1]
|
||||
x = x + d * dt
|
||||
|
||||
return x / sigmas[-1]
|
||||
|
||||
def generate(self, prompt: str, init_img=None, init_mask=None, mask_mode='mask', resize_mode='resize', noise_mode='seed',
|
||||
denoising_strength:float=0.8, ddim_steps=50, sampler_name='k_lms', n_iter=1, batch_size=1, cfg_scale=7.5, seed=None,
|
||||
height=512, width=512, save_individual_images: bool = True, save_grid: bool = True, ddim_eta:float = 0.0):
|
||||
seed = seed_to_int(seed)
|
||||
image_dict = {
|
||||
"seed": seed
|
||||
}
|
||||
# Init image is assumed to be a PIL image
|
||||
init_img = self.resize_image('resize', init_img, width, height)
|
||||
if sampler_name == 'PLMS':
|
||||
sampler = PLMSSampler(self.model)
|
||||
elif sampler_name == 'DDIM':
|
||||
sampler = DDIMSampler(self.model)
|
||||
elif sampler_name == 'k_dpm_2_a':
|
||||
sampler = KDiffusionSampler(self.model,'dpm_2_ancestral')
|
||||
elif sampler_name == 'k_dpm_2':
|
||||
sampler = KDiffusionSampler(self.model,'dpm_2')
|
||||
elif sampler_name == 'k_euler_a':
|
||||
sampler = KDiffusionSampler(self.model,'euler_ancestral')
|
||||
elif sampler_name == 'k_euler':
|
||||
sampler = KDiffusionSampler(self.model,'euler')
|
||||
elif sampler_name == 'k_heun':
|
||||
sampler = KDiffusionSampler(self.model,'heun')
|
||||
elif sampler_name == 'k_lms':
|
||||
sampler = KDiffusionSampler(self.model,'lms')
|
||||
else:
|
||||
raise Exception("Unknown sampler: " + sampler_name)
|
||||
|
||||
torch_gc()
|
||||
def process_init_mask(init_mask: PIL.Image):
|
||||
if init_mask.mode == "RGBA":
|
||||
init_mask = init_mask.convert('RGBA')
|
||||
background = PIL.Image.new('RGBA', init_mask.size, (0, 0, 0))
|
||||
init_mask = PIL.Image.alpha_composite(background, init_mask)
|
||||
init_mask = init_mask.convert('RGB')
|
||||
return init_mask
|
||||
|
||||
if mask_mode == "mask":
|
||||
if init_mask:
|
||||
init_mask = process_init_mask(init_mask)
|
||||
elif mask_mode == "invert":
|
||||
if init_mask:
|
||||
init_mask = process_init_mask(init_mask)
|
||||
init_mask = PIL.ImageOps.invert(init_mask)
|
||||
elif mask_mode == "alpha":
|
||||
init_img_transparency = init_img.split()[-1].convert('L')#.point(lambda x: 255 if x > 0 else 0, mode='1')
|
||||
init_mask = init_img_transparency
|
||||
init_mask = init_mask.convert("RGB")
|
||||
init_mask = self.resize_image(resize_mode, init_mask, width, height)
|
||||
init_mask = init_mask.convert("RGB")
|
||||
|
||||
assert 0. <= denoising_strength <= 1., 'can only work with strength in [0.0, 1.0]'
|
||||
t_enc = int(denoising_strength * ddim_steps)
|
||||
|
||||
if init_mask is not None and (noise_mode == "matched" or noise_mode == "find_and_matched") and init_img is not None:
|
||||
noise_q = 0.99
|
||||
color_variation = 0.0
|
||||
mask_blend_factor = 1.0
|
||||
|
||||
np_init = (np.asarray(init_img.convert("RGB"))/255.0).astype(np.float64) # annoyingly complex mask fixing
|
||||
np_mask_rgb = 1. - (np.asarray(PIL.ImageOps.invert(init_mask).convert("RGB"))/255.0).astype(np.float64)
|
||||
np_mask_rgb -= np.min(np_mask_rgb)
|
||||
np_mask_rgb /= np.max(np_mask_rgb)
|
||||
np_mask_rgb = 1. - np_mask_rgb
|
||||
np_mask_rgb_hardened = 1. - (np_mask_rgb < 0.99).astype(np.float64)
|
||||
blurred = skimage.filters.gaussian(np_mask_rgb_hardened[:], sigma=16., channel_axis=2, truncate=32.)
|
||||
blurred2 = skimage.filters.gaussian(np_mask_rgb_hardened[:], sigma=16., channel_axis=2, truncate=32.)
|
||||
#np_mask_rgb_dilated = np_mask_rgb + blurred # fixup mask todo: derive magic constants
|
||||
#np_mask_rgb = np_mask_rgb + blurred
|
||||
np_mask_rgb_dilated = np.clip((np_mask_rgb + blurred2) * 0.7071, 0., 1.)
|
||||
np_mask_rgb = np.clip((np_mask_rgb + blurred) * 0.7071, 0., 1.)
|
||||
|
||||
noise_rgb = self.get_matched_noise(np_init, np_mask_rgb, noise_q, color_variation)
|
||||
blend_mask_rgb = np.clip(np_mask_rgb_dilated,0.,1.) ** (mask_blend_factor)
|
||||
noised = noise_rgb[:]
|
||||
blend_mask_rgb **= (2.)
|
||||
noised = np_init[:] * (1. - blend_mask_rgb) + noised * blend_mask_rgb
|
||||
|
||||
np_mask_grey = np.sum(np_mask_rgb, axis=2)/3.
|
||||
ref_mask = np_mask_grey < 1e-3
|
||||
|
||||
all_mask = np.ones((height, width), dtype=bool)
|
||||
noised[all_mask,:] = skimage.exposure.match_histograms(noised[all_mask,:]**1., noised[ref_mask,:], channel_axis=1)
|
||||
|
||||
init_img = PIL.Image.fromarray(np.clip(noised * 255., 0., 255.).astype(np.uint8), mode="RGB")
|
||||
|
||||
def init():
|
||||
image = init_img.convert('RGB')
|
||||
image = np.array(image).astype(np.float32) / 255.0
|
||||
image = image[None].transpose(0, 3, 1, 2)
|
||||
image = torch.from_numpy(image)
|
||||
|
||||
mask_channel = None
|
||||
if init_mask:
|
||||
alpha = self.resize_image(resize_mode, init_mask, width // 8, height // 8)
|
||||
mask_channel = alpha.split()[-1]
|
||||
|
||||
mask = None
|
||||
if mask_channel is not None:
|
||||
mask = np.array(mask_channel).astype(np.float32) / 255.0
|
||||
mask = (1 - mask)
|
||||
mask = np.tile(mask, (4, 1, 1))
|
||||
mask = mask[None].transpose(0, 1, 2, 3)
|
||||
mask = torch.from_numpy(mask).to(self.model.device)
|
||||
|
||||
init_image = 2. * image - 1.
|
||||
init_image = init_image.to(self.model.device)
|
||||
init_latent = self.model.get_first_stage_encoding(self.model.encode_first_stage(init_image)) # move to latent space
|
||||
|
||||
return init_latent, mask,
|
||||
|
||||
def sample(init_data, x, conditioning, unconditional_conditioning, sampler_name):
|
||||
t_enc_steps = t_enc
|
||||
obliterate = False
|
||||
if ddim_steps == t_enc_steps:
|
||||
t_enc_steps = t_enc_steps - 1
|
||||
obliterate = True
|
||||
|
||||
if sampler_name != 'DDIM':
|
||||
x0, z_mask = init_data
|
||||
|
||||
sigmas = sampler.model_wrap.get_sigmas(ddim_steps)
|
||||
noise = x * sigmas[ddim_steps - t_enc_steps - 1]
|
||||
|
||||
xi = x0 + noise
|
||||
|
||||
# Obliterate masked image
|
||||
if z_mask is not None and obliterate:
|
||||
random = torch.randn(z_mask.shape, device=xi.device)
|
||||
xi = (z_mask * noise) + ((1-z_mask) * xi)
|
||||
|
||||
sigma_sched = sigmas[ddim_steps - t_enc_steps - 1:]
|
||||
model_wrap_cfg = CFGMaskedDenoiser(sampler.model_wrap)
|
||||
samples_ddim = K.sampling.__dict__[f'sample_{sampler.get_sampler_name()}'](model_wrap_cfg, xi, sigma_sched,
|
||||
extra_args={'cond': conditioning, 'uncond': unconditional_conditioning,
|
||||
'cond_scale': cfg_scale, 'mask': z_mask, 'x0': x0, 'xi': xi}, disable=False)
|
||||
else:
|
||||
|
||||
x0, z_mask = init_data
|
||||
|
||||
sampler.make_schedule(ddim_num_steps=ddim_steps, ddim_eta=0.0, verbose=False)
|
||||
z_enc = sampler.stochastic_encode(x0, torch.tensor([t_enc_steps]*batch_size).to(self.model.device))
|
||||
|
||||
# Obliterate masked image
|
||||
if z_mask is not None and obliterate:
|
||||
random = torch.randn(z_mask.shape, device=z_enc.device)
|
||||
z_enc = (z_mask * random) + ((1-z_mask) * z_enc)
|
||||
|
||||
# decode it
|
||||
samples_ddim = sampler.decode(z_enc, conditioning, t_enc_steps,
|
||||
unconditional_guidance_scale=cfg_scale,
|
||||
unconditional_conditioning=unconditional_conditioning,
|
||||
z_mask=z_mask, x0=x0)
|
||||
return samples_ddim
|
||||
|
||||
torch_gc()
|
||||
|
||||
if self.load_concepts and self.concepts_dir is not None:
|
||||
prompt_tokens = re.findall('<([a-zA-Z0-9-]+)>', prompt)
|
||||
if prompt_tokens:
|
||||
self.process_prompt_tokens(prompt_tokens)
|
||||
|
||||
os.makedirs(self.output_dir, exist_ok=True)
|
||||
|
||||
sample_path = os.path.join(self.output_dir, "samples")
|
||||
os.makedirs(sample_path, exist_ok=True)
|
||||
|
||||
if self.verify_input:
|
||||
try:
|
||||
check_prompt_length(self.model, prompt, self.comments)
|
||||
except:
|
||||
import traceback
|
||||
print("Error verifying input:", file=sys.stderr)
|
||||
print(traceback.format_exc(), file=sys.stderr)
|
||||
|
||||
all_prompts = batch_size * n_iter * [prompt]
|
||||
all_seeds = [seed + x for x in range(len(all_prompts))]
|
||||
|
||||
precision_scope = torch.autocast if self.auto_cast else nullcontext
|
||||
|
||||
with torch.no_grad(), precision_scope("cuda"):
|
||||
for n in range(n_iter):
|
||||
print(f"Iteration: {n+1}/{n_iter}")
|
||||
prompts = all_prompts[n * batch_size:(n + 1) * batch_size]
|
||||
seeds = all_seeds[n * batch_size:(n + 1) * batch_size]
|
||||
|
||||
uc = self.model.get_learned_conditioning(len(prompts) * [''])
|
||||
|
||||
if isinstance(prompts, tuple):
|
||||
prompts = list(prompts)
|
||||
|
||||
c = self.model.get_learned_conditioning(prompts)
|
||||
|
||||
opt_C = 4
|
||||
opt_f = 8
|
||||
shape = [opt_C, height // opt_f, width // opt_f]
|
||||
|
||||
x = self.create_random_tensors(shape, seeds=seeds)
|
||||
init_data = init()
|
||||
samples_ddim = sample(init_data=init_data, x=x, conditioning=c, unconditional_conditioning=uc, sampler_name=sampler_name)
|
||||
|
||||
x_samples_ddim = self.model.decode_first_stage(samples_ddim)
|
||||
x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
|
||||
|
||||
for i, x_sample in enumerate(x_samples_ddim):
|
||||
sanitized_prompt = slugify(prompts[i])
|
||||
full_path = os.path.join(os.getcwd(), sample_path)
|
||||
sample_path_i = sample_path
|
||||
base_count = get_next_sequence_number(sample_path_i)
|
||||
filename = f"{base_count:05}-{ddim_steps}_{sampler_name}_{seeds[i]}_{sanitized_prompt}"[:200-len(full_path)]
|
||||
|
||||
x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
|
||||
x_sample = x_sample.astype(np.uint8)
|
||||
image = PIL.Image.fromarray(x_sample)
|
||||
image_dict['image'] = image
|
||||
self.images.append(image_dict)
|
||||
|
||||
if save_individual_images:
|
||||
path = os.path.join(sample_path, filename + '.' + self.save_extension)
|
||||
success = save_sample(image, filename, sample_path_i, self.save_extension)
|
||||
if success:
|
||||
if self.output_file_path:
|
||||
self.output_images.append(path)
|
||||
else:
|
||||
self.output_images.append(image)
|
||||
else:
|
||||
return
|
||||
|
||||
self.info = f"""
|
||||
{prompt}
|
||||
Steps: {ddim_steps}, Sampler: {sampler_name}, CFG scale: {cfg_scale}, Seed: {seed}
|
||||
""".strip()
|
||||
self.stats = f'''
|
||||
'''
|
||||
|
||||
for comment in self.comments:
|
||||
self.info += "\n\n" + comment
|
||||
|
||||
torch_gc()
|
||||
|
||||
del sampler
|
||||
|
||||
return
|
||||
@ -1,201 +0,0 @@
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
from contextlib import contextmanager, nullcontext
|
||||
|
||||
import numpy as np
|
||||
import PIL
|
||||
import torch
|
||||
from einops import rearrange
|
||||
from ldm.models.diffusion.ddim import DDIMSampler
|
||||
from ldm.models.diffusion.kdiffusion import KDiffusionSampler
|
||||
from ldm.models.diffusion.plms import PLMSSampler
|
||||
from nataili.util.cache import torch_gc
|
||||
from nataili.util.check_prompt_length import check_prompt_length
|
||||
from nataili.util.get_next_sequence_number import get_next_sequence_number
|
||||
from nataili.util.image_grid import image_grid
|
||||
from nataili.util.load_learned_embed_in_clip import load_learned_embed_in_clip
|
||||
from nataili.util.save_sample import save_sample
|
||||
from nataili.util.seed_to_int import seed_to_int
|
||||
from slugify import slugify
|
||||
|
||||
|
||||
class txt2img:
|
||||
def __init__(self, model, device, output_dir, save_extension='jpg',
|
||||
output_file_path=False, load_concepts=False, concepts_dir=None,
|
||||
verify_input=True, auto_cast=True):
|
||||
self.model = model
|
||||
self.output_dir = output_dir
|
||||
self.output_file_path = output_file_path
|
||||
self.save_extension = save_extension
|
||||
self.load_concepts = load_concepts
|
||||
self.concepts_dir = concepts_dir
|
||||
self.verify_input = verify_input
|
||||
self.auto_cast = auto_cast
|
||||
self.device = device
|
||||
self.comments = []
|
||||
self.output_images = []
|
||||
self.info = ''
|
||||
self.stats = ''
|
||||
self.images = []
|
||||
|
||||
def create_random_tensors(self, shape, seeds):
|
||||
xs = []
|
||||
for seed in seeds:
|
||||
torch.manual_seed(seed)
|
||||
|
||||
# randn results depend on device; gpu and cpu get different results for same seed;
|
||||
# the way I see it, it's better to do this on CPU, so that everyone gets same result;
|
||||
# but the original script had it like this so i do not dare change it for now because
|
||||
# it will break everyone's seeds.
|
||||
xs.append(torch.randn(shape, device=self.device))
|
||||
x = torch.stack(xs)
|
||||
return x
|
||||
|
||||
def process_prompt_tokens(self, prompt_tokens):
|
||||
# compviz codebase
|
||||
tokenizer = self.model.cond_stage_model.tokenizer
|
||||
text_encoder = self.model.cond_stage_model.transformer
|
||||
|
||||
# diffusers codebase
|
||||
#tokenizer = pipe.tokenizer
|
||||
#text_encoder = pipe.text_encoder
|
||||
|
||||
ext = ('.pt', '.bin')
|
||||
for token_name in prompt_tokens:
|
||||
embedding_path = os.path.join(self.concepts_dir, token_name)
|
||||
if os.path.exists(embedding_path):
|
||||
for files in os.listdir(embedding_path):
|
||||
if files.endswith(ext):
|
||||
load_learned_embed_in_clip(f"{os.path.join(embedding_path, files)}", text_encoder, tokenizer, f"<{token_name}>")
|
||||
else:
|
||||
print(f"Concept {token_name} not found in {self.concepts_dir}")
|
||||
del tokenizer, text_encoder
|
||||
return
|
||||
del tokenizer, text_encoder
|
||||
|
||||
def generate(self, prompt: str, ddim_steps=50, sampler_name='k_lms', n_iter=1, batch_size=1, cfg_scale=7.5, seed=None,
|
||||
height=512, width=512, save_individual_images: bool = True, save_grid: bool = True, ddim_eta:float = 0.0):
|
||||
seed = seed_to_int(seed)
|
||||
|
||||
image_dict = {
|
||||
"seed": seed
|
||||
}
|
||||
negprompt = ''
|
||||
if '###' in prompt:
|
||||
prompt, negprompt = prompt.split('###', 1)
|
||||
prompt = prompt.strip()
|
||||
negprompt = negprompt.strip()
|
||||
|
||||
if sampler_name == 'PLMS':
|
||||
sampler = PLMSSampler(self.model)
|
||||
elif sampler_name == 'DDIM':
|
||||
sampler = DDIMSampler(self.model)
|
||||
elif sampler_name == 'k_dpm_2_a':
|
||||
sampler = KDiffusionSampler(self.model,'dpm_2_ancestral')
|
||||
elif sampler_name == 'k_dpm_2':
|
||||
sampler = KDiffusionSampler(self.model,'dpm_2')
|
||||
elif sampler_name == 'k_euler_a':
|
||||
sampler = KDiffusionSampler(self.model,'euler_ancestral')
|
||||
elif sampler_name == 'k_euler':
|
||||
sampler = KDiffusionSampler(self.model,'euler')
|
||||
elif sampler_name == 'k_heun':
|
||||
sampler = KDiffusionSampler(self.model,'heun')
|
||||
elif sampler_name == 'k_lms':
|
||||
sampler = KDiffusionSampler(self.model,'lms')
|
||||
else:
|
||||
raise Exception("Unknown sampler: " + sampler_name)
|
||||
|
||||
def sample(init_data, x, conditioning, unconditional_conditioning, sampler_name):
|
||||
samples_ddim, _ = sampler.sample(S=ddim_steps, conditioning=conditioning, unconditional_guidance_scale=cfg_scale,
|
||||
unconditional_conditioning=unconditional_conditioning, x_T=x)
|
||||
return samples_ddim
|
||||
|
||||
torch_gc()
|
||||
|
||||
if self.load_concepts and self.concepts_dir is not None:
|
||||
prompt_tokens = re.findall('<([a-zA-Z0-9-]+)>', prompt)
|
||||
if prompt_tokens:
|
||||
self.process_prompt_tokens(prompt_tokens)
|
||||
|
||||
os.makedirs(self.output_dir, exist_ok=True)
|
||||
|
||||
sample_path = os.path.join(self.output_dir, "samples")
|
||||
os.makedirs(sample_path, exist_ok=True)
|
||||
|
||||
if self.verify_input:
|
||||
try:
|
||||
check_prompt_length(self.model, prompt, self.comments)
|
||||
except:
|
||||
import traceback
|
||||
print("Error verifying input:", file=sys.stderr)
|
||||
print(traceback.format_exc(), file=sys.stderr)
|
||||
|
||||
all_prompts = batch_size * n_iter * [prompt]
|
||||
all_seeds = [seed + x for x in range(len(all_prompts))]
|
||||
|
||||
precision_scope = torch.autocast if self.auto_cast else nullcontext
|
||||
|
||||
with torch.no_grad(), precision_scope("cuda"):
|
||||
for n in range(n_iter):
|
||||
print(f"Iteration: {n+1}/{n_iter}")
|
||||
prompts = all_prompts[n * batch_size:(n + 1) * batch_size]
|
||||
seeds = all_seeds[n * batch_size:(n + 1) * batch_size]
|
||||
|
||||
uc = self.model.get_learned_conditioning(len(prompts) * [negprompt])
|
||||
|
||||
if isinstance(prompts, tuple):
|
||||
prompts = list(prompts)
|
||||
|
||||
c = self.model.get_learned_conditioning(prompts)
|
||||
|
||||
opt_C = 4
|
||||
opt_f = 8
|
||||
shape = [opt_C, height // opt_f, width // opt_f]
|
||||
|
||||
x = self.create_random_tensors(shape, seeds=seeds)
|
||||
|
||||
samples_ddim = sample(init_data=None, x=x, conditioning=c, unconditional_conditioning=uc, sampler_name=sampler_name)
|
||||
|
||||
x_samples_ddim = self.model.decode_first_stage(samples_ddim)
|
||||
x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
|
||||
|
||||
for i, x_sample in enumerate(x_samples_ddim):
|
||||
sanitized_prompt = slugify(prompts[i])
|
||||
full_path = os.path.join(os.getcwd(), sample_path)
|
||||
sample_path_i = sample_path
|
||||
base_count = get_next_sequence_number(sample_path_i)
|
||||
filename = f"{base_count:05}-{ddim_steps}_{sampler_name}_{seeds[i]}_{sanitized_prompt}"[:200-len(full_path)]
|
||||
|
||||
x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
|
||||
x_sample = x_sample.astype(np.uint8)
|
||||
image = PIL.Image.fromarray(x_sample)
|
||||
image_dict['image'] = image
|
||||
self.images.append(image_dict)
|
||||
|
||||
if save_individual_images:
|
||||
path = os.path.join(sample_path, filename + '.' + self.save_extension)
|
||||
success = save_sample(image, filename, sample_path_i, self.save_extension)
|
||||
if success:
|
||||
if self.output_file_path:
|
||||
self.output_images.append(path)
|
||||
else:
|
||||
self.output_images.append(image)
|
||||
else:
|
||||
return
|
||||
|
||||
self.info = f"""
|
||||
{prompt}
|
||||
Steps: {ddim_steps}, Sampler: {sampler_name}, CFG scale: {cfg_scale}, Seed: {seed}
|
||||
""".strip()
|
||||
self.stats = f'''
|
||||
'''
|
||||
|
||||
for comment in self.comments:
|
||||
self.info += "\n\n" + comment
|
||||
|
||||
torch_gc()
|
||||
|
||||
del sampler
|
||||
|
||||
return
|
||||
@ -1,458 +0,0 @@
|
||||
import os
|
||||
import json
|
||||
import shutil
|
||||
import zipfile
|
||||
import requests
|
||||
import git
|
||||
import torch
|
||||
import hashlib
|
||||
from ldm.util import instantiate_from_config
|
||||
from omegaconf import OmegaConf
|
||||
from transformers import logging
|
||||
|
||||
from basicsr.archs.rrdbnet_arch import RRDBNet
|
||||
from gfpgan import GFPGANer
|
||||
from realesrgan import RealESRGANer
|
||||
from ldm.models.blip import blip_decoder
|
||||
from tqdm import tqdm
|
||||
import open_clip
|
||||
import clip
|
||||
|
||||
from nataili.util.cache import torch_gc
|
||||
from nataili.util import logger
|
||||
|
||||
logging.set_verbosity_error()
|
||||
|
||||
models = json.load(open('./db.json'))
|
||||
dependencies = json.load(open('./db_dep.json'))
|
||||
remote_models = "https://raw.githubusercontent.com/Sygil-Dev/nataili-model-reference/main/db.json"
|
||||
remote_dependencies = "https://raw.githubusercontent.com/Sygil-Dev/nataili-model-reference/main/db_dep.json"
|
||||
|
||||
class ModelManager():
|
||||
def __init__(self, hf_auth=None, download=True):
|
||||
if download:
|
||||
try:
|
||||
logger.init("Model Reference", status="Downloading")
|
||||
r = requests.get(remote_models)
|
||||
self.models = r.json()
|
||||
r = requests.get(remote_dependencies)
|
||||
self.dependencies = json.load(open('./db_dep.json'))
|
||||
logger.init_ok("Model Reference", status="OK")
|
||||
except:
|
||||
logger.init_err("Model Reference", status="Download Error")
|
||||
self.models = json.load(open('./db.json'))
|
||||
self.dependencies = json.load(open('./db_dep.json'))
|
||||
logger.init_warn("Model Reference", status="Local")
|
||||
self.available_models = []
|
||||
self.tainted_models = []
|
||||
self.available_dependencies = []
|
||||
self.loaded_models = {}
|
||||
self.hf_auth = None
|
||||
self.set_authentication(hf_auth)
|
||||
|
||||
def init(self):
|
||||
dependencies_available = []
|
||||
for dependency in self.dependencies:
|
||||
if self.check_available(self.get_dependency_files(dependency)):
|
||||
dependencies_available.append(dependency)
|
||||
self.available_dependencies = dependencies_available
|
||||
|
||||
models_available = []
|
||||
for model in self.models:
|
||||
if self.check_available(self.get_model_files(model)):
|
||||
models_available.append(model)
|
||||
self.available_models = models_available
|
||||
|
||||
if self.hf_auth is not None:
|
||||
if 'username' not in self.hf_auth and 'password' not in self.hf_auth:
|
||||
raise ValueError('hf_auth must contain username and password')
|
||||
else:
|
||||
if self.hf_auth['username'] == '' or self.hf_auth['password'] == '':
|
||||
raise ValueError('hf_auth must contain username and password')
|
||||
return True
|
||||
|
||||
def set_authentication(self, hf_auth=None):
|
||||
# We do not let No authentication override previously set auth
|
||||
if not hf_auth and self.hf_auth:
|
||||
return
|
||||
self.hf_auth = hf_auth
|
||||
|
||||
def get_model(self, model_name):
|
||||
return self.models.get(model_name)
|
||||
|
||||
def get_filtered_models(self, **kwargs):
|
||||
'''Get all model names.
|
||||
Can filter based on metadata of the model reference db
|
||||
'''
|
||||
filtered_models = self.models
|
||||
for keyword in kwargs:
|
||||
iterating_models = filtered_models.copy()
|
||||
filtered_models = {}
|
||||
for model in iterating_models:
|
||||
# logger.debug([keyword,iterating_models[model].get(keyword),kwargs[keyword]])
|
||||
if iterating_models[model].get(keyword) == kwargs[keyword]:
|
||||
filtered_models[model] = iterating_models[model]
|
||||
return filtered_models
|
||||
|
||||
def get_filtered_model_names(self, **kwargs):
|
||||
filtered_models = self.get_filtered_models(**kwargs)
|
||||
return list(filtered_models.keys())
|
||||
|
||||
def get_dependency(self, dependency_name):
|
||||
return self.dependencies[dependency_name]
|
||||
|
||||
def get_model_files(self, model_name):
|
||||
return self.models[model_name]['config']['files']
|
||||
|
||||
def get_dependency_files(self, dependency_name):
|
||||
return self.dependencies[dependency_name]['config']['files']
|
||||
|
||||
def get_model_download(self, model_name):
|
||||
return self.models[model_name]['config']['download']
|
||||
|
||||
def get_dependency_download(self, dependency_name):
|
||||
return self.dependencies[dependency_name]['config']['download']
|
||||
|
||||
def get_available_models(self):
|
||||
return self.available_models
|
||||
|
||||
def get_available_dependencies(self):
|
||||
return self.available_dependencies
|
||||
|
||||
def get_loaded_models(self):
|
||||
return self.loaded_models
|
||||
|
||||
def get_loaded_models_names(self):
|
||||
return list(self.loaded_models.keys())
|
||||
|
||||
def get_loaded_model(self, model_name):
|
||||
return self.loaded_models[model_name]
|
||||
|
||||
def unload_model(self, model_name):
|
||||
if model_name in self.loaded_models:
|
||||
del self.loaded_models[model_name]
|
||||
return True
|
||||
return False
|
||||
|
||||
def unload_all_models(self):
|
||||
for model in self.loaded_models:
|
||||
del self.loaded_models[model]
|
||||
return True
|
||||
|
||||
def taint_model(self,model_name):
|
||||
'''Marks a model as not valid by remiving it from available_models'''
|
||||
if model_name in self.available_models:
|
||||
self.available_models.remove(model_name)
|
||||
self.tainted_models.append(model_name)
|
||||
|
||||
def taint_models(self, models):
|
||||
for model in models:
|
||||
self.taint_model(model)
|
||||
|
||||
def load_model_from_config(self, model_path='', config_path='', map_location="cpu"):
|
||||
config = OmegaConf.load(config_path)
|
||||
pl_sd = torch.load(model_path, map_location=map_location)
|
||||
if "global_step" in pl_sd:
|
||||
logger.info(f"Global Step: {pl_sd['global_step']}")
|
||||
sd = pl_sd["state_dict"]
|
||||
model = instantiate_from_config(config.model)
|
||||
m, u = model.load_state_dict(sd, strict=False)
|
||||
model = model.eval()
|
||||
del pl_sd, sd, m, u
|
||||
return model
|
||||
|
||||
def load_ckpt(self, model_name='', precision='half', gpu_id=0):
|
||||
ckpt_path = self.get_model_files(model_name)[0]['path']
|
||||
config_path = self.get_model_files(model_name)[1]['path']
|
||||
model = self.load_model_from_config(model_path=ckpt_path, config_path=config_path)
|
||||
device = torch.device(f"cuda:{gpu_id}")
|
||||
model = (model if precision=='full' else model.half()).to(device)
|
||||
torch_gc()
|
||||
return {'model': model, 'device': device}
|
||||
|
||||
def load_realesrgan(self, model_name='', precision='half', gpu_id=0):
|
||||
|
||||
RealESRGAN_models = {
|
||||
'RealESRGAN_x4plus': RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4),
|
||||
'RealESRGAN_x4plus_anime_6B': RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=6, num_grow_ch=32, scale=4)
|
||||
}
|
||||
|
||||
model_path = self.get_model_files(model_name)[0]['path']
|
||||
device = torch.device(f"cuda:{gpu_id}")
|
||||
model = RealESRGANer(scale=2, model_path=model_path, model=RealESRGAN_models[models[model_name]['name']],
|
||||
pre_pad=0, half=True if precision == 'half' else False, device=device)
|
||||
return {'model': model, 'device': device}
|
||||
|
||||
def load_gfpgan(self, model_name='', gpu_id=0):
|
||||
|
||||
model_path = self.get_model_files(model_name)[0]['path']
|
||||
device = torch.device(f"cuda:{gpu_id}")
|
||||
model = GFPGANer(model_path=model_path, upscale=1, arch='clean',
|
||||
channel_multiplier=2, bg_upsampler=None, device=device)
|
||||
return {'model': model, 'device': device}
|
||||
|
||||
def load_blip(self, model_name='', precision='half', gpu_id=0, blip_image_eval_size=512, vit='base'):
|
||||
# vit = 'base' or 'large'
|
||||
model_path = self.get_model_files(model_name)[0]['path']
|
||||
device = torch.device(f"cuda:{gpu_id}")
|
||||
model = blip_decoder(pretrained=model_path,
|
||||
med_config="configs/blip/med_config.json",
|
||||
image_size=blip_image_eval_size, vit=vit)
|
||||
model = model.eval()
|
||||
model = (model if precision=='full' else model.half()).to(device)
|
||||
return {'model': model, 'device': device}
|
||||
|
||||
def load_open_clip(self, model_name='', precision='half', gpu_id=0):
|
||||
pretrained = self.get_model(model_name)['pretrained_name']
|
||||
device = torch.device(f"cuda:{gpu_id}")
|
||||
model, _, preprocesses = open_clip.create_model_and_transforms(model_name, pretrained=pretrained, cache_dir='models/clip')
|
||||
model = model.eval()
|
||||
model = (model if precision=='full' else model.half()).to(device)
|
||||
return {'model': model, 'device': device, 'preprocesses': preprocesses}
|
||||
|
||||
def load_clip(self, model_name='', precision='half', gpu_id=0):
|
||||
device = torch.device(f"cuda:{gpu_id}")
|
||||
model, preprocesses = clip.load(model_name, device=device, download_root='models/clip')
|
||||
model = model.eval()
|
||||
model = (model if precision=='full' else model.half()).to(device)
|
||||
return {'model': model, 'device': device, 'preprocesses': preprocesses}
|
||||
|
||||
def load_model(self, model_name='', precision='half', gpu_id=0):
|
||||
if model_name not in self.available_models:
|
||||
return False
|
||||
if self.models[model_name]['type'] == 'ckpt':
|
||||
self.loaded_models[model_name] = self.load_ckpt(model_name, precision, gpu_id)
|
||||
return True
|
||||
elif self.models[model_name]['type'] == 'realesrgan':
|
||||
self.loaded_models[model_name] = self.load_realesrgan(model_name, precision, gpu_id)
|
||||
return True
|
||||
elif self.models[model_name]['type'] == 'gfpgan':
|
||||
self.loaded_models[model_name] = self.load_gfpgan(model_name, gpu_id)
|
||||
return True
|
||||
elif self.models[model_name]['type'] == 'blip':
|
||||
self.loaded_models[model_name] = self.load_blip(model_name, precision, gpu_id, 512, 'base')
|
||||
return True
|
||||
elif self.models[model_name]['type'] == 'open_clip':
|
||||
self.loaded_models[model_name] = self.load_open_clip(model_name, precision, gpu_id)
|
||||
return True
|
||||
elif self.models[model_name]['type'] == 'clip':
|
||||
self.loaded_models[model_name] = self.load_clip(model_name, precision, gpu_id)
|
||||
return True
|
||||
else:
|
||||
return False
|
||||
|
||||
def validate_model(self, model_name):
|
||||
files = self.get_model_files(model_name)
|
||||
all_ok = True
|
||||
for file_details in files:
|
||||
if not self.check_file_available(file_details['path']):
|
||||
return False
|
||||
if not self.validate_file(file_details):
|
||||
return False
|
||||
return True
|
||||
|
||||
def validate_file(self, file_details):
|
||||
if 'md5sum' in file_details:
|
||||
file_name = file_details['path']
|
||||
logger.debug(f"Getting md5sum of {file_name}")
|
||||
with open(file_name, 'rb') as file_to_check:
|
||||
file_hash = hashlib.md5()
|
||||
while chunk := file_to_check.read(8192):
|
||||
file_hash.update(chunk)
|
||||
if file_details['md5sum'] != file_hash.hexdigest():
|
||||
return False
|
||||
return True
|
||||
|
||||
def check_file_available(self, file_path):
|
||||
return os.path.exists(file_path)
|
||||
|
||||
def check_available(self, files):
|
||||
available = True
|
||||
for file in files:
|
||||
if not self.check_file_available(file['path']):
|
||||
available = False
|
||||
return available
|
||||
|
||||
def download_file(self, url, file_path):
|
||||
# make directory
|
||||
os.makedirs(os.path.dirname(file_path), exist_ok=True)
|
||||
pbar_desc = file_path.split('/')[-1]
|
||||
r = requests.get(url, stream=True)
|
||||
with open(file_path, 'wb') as f:
|
||||
with tqdm(
|
||||
# all optional kwargs
|
||||
unit='B', unit_scale=True, unit_divisor=1024, miniters=1,
|
||||
desc=pbar_desc, total=int(r.headers.get('content-length', 0))
|
||||
) as pbar:
|
||||
for chunk in r.iter_content(chunk_size=16*1024):
|
||||
if chunk:
|
||||
f.write(chunk)
|
||||
pbar.update(len(chunk))
|
||||
|
||||
def download_model(self, model_name):
|
||||
if model_name in self.available_models:
|
||||
logger.info(f"{model_name} is already available.")
|
||||
return True
|
||||
download = self.get_model_download(model_name)
|
||||
files = self.get_model_files(model_name)
|
||||
for i in range(len(download)):
|
||||
file_path = f"{download[i]['file_path']}/{download[i]['file_name']}" if 'file_path' in download[i] else files[i]['path']
|
||||
|
||||
if 'file_url' in download[i]:
|
||||
download_url = download[i]['file_url']
|
||||
if 'hf_auth' in download[i]:
|
||||
username = self.hf_auth['username']
|
||||
password = self.hf_auth['password']
|
||||
download_url = download_url.format(username=username, password=password)
|
||||
if 'file_name' in download[i]:
|
||||
download_name = download[i]['file_name']
|
||||
if 'file_path' in download[i]:
|
||||
download_path = download[i]['file_path']
|
||||
|
||||
if 'manual' in download[i]:
|
||||
logger.warning(f"The model {model_name} requires manual download from {download_url}. Please place it in {download_path}/{download_name} then press ENTER to continue...")
|
||||
input('')
|
||||
continue
|
||||
# TODO: simplify
|
||||
if "file_content" in download[i]:
|
||||
file_content = download[i]['file_content']
|
||||
logger.info(f"writing {file_content} to {file_path}")
|
||||
# make directory download_path
|
||||
os.makedirs(download_path, exist_ok=True)
|
||||
# write file_content to download_path/download_name
|
||||
with open(os.path.join(download_path, download_name), 'w') as f:
|
||||
f.write(file_content)
|
||||
elif 'symlink' in download[i]:
|
||||
logger.info(f"symlink {file_path} to {download[i]['symlink']}")
|
||||
symlink = download[i]['symlink']
|
||||
# make directory symlink
|
||||
os.makedirs(download_path, exist_ok=True)
|
||||
# make symlink from download_path/download_name to symlink
|
||||
os.symlink(symlink, os.path.join(download_path, download_name))
|
||||
elif 'git' in download[i]:
|
||||
logger.info(f"git clone {download_url} to {file_path}")
|
||||
# make directory download_path
|
||||
os.makedirs(file_path, exist_ok=True)
|
||||
git.Git(file_path).clone(download_url)
|
||||
if 'post_process' in download[i]:
|
||||
for post_process in download[i]['post_process']:
|
||||
if 'delete' in post_process:
|
||||
# delete folder post_process['delete']
|
||||
logger.info(f"delete {post_process['delete']}")
|
||||
try:
|
||||
shutil.rmtree(post_process['delete'])
|
||||
except PermissionError as e:
|
||||
logger.error(f"[!] Something went wrong while deleting the `{post_process['delete']}`. Please delete it manually.")
|
||||
logger.error("PermissionError: ", e)
|
||||
else:
|
||||
if not self.check_file_available(file_path) or model_name in self.tainted_models:
|
||||
logger.debug(f'Downloading {download_url} to {file_path}')
|
||||
self.download_file(download_url, file_path)
|
||||
if not self.validate_model(model_name):
|
||||
return False
|
||||
if model_name in self.tainted_models:
|
||||
self.tainted_models.remove(model_name)
|
||||
self.init()
|
||||
return True
|
||||
|
||||
def download_dependency(self, dependency_name):
|
||||
if dependency_name in self.available_dependencies:
|
||||
logger.info(f"{dependency_name} is already installed.")
|
||||
return True
|
||||
download = self.get_dependency_download(dependency_name)
|
||||
files = self.get_dependency_files(dependency_name)
|
||||
for i in range(len(download)):
|
||||
if "git" in download[i]:
|
||||
logger.warning("git download not implemented yet")
|
||||
break
|
||||
|
||||
file_path = files[i]['path']
|
||||
if 'file_url' in download[i]:
|
||||
download_url = download[i]['file_url']
|
||||
if 'file_name' in download[i]:
|
||||
download_name = download[i]['file_name']
|
||||
if 'file_path' in download[i]:
|
||||
download_path = download[i]['file_path']
|
||||
logger.debug(download_name)
|
||||
if "unzip" in download[i]:
|
||||
zip_path = f'temp/{download_name}.zip'
|
||||
# os dirname zip_path
|
||||
# mkdir temp
|
||||
os.makedirs("temp", exist_ok=True)
|
||||
|
||||
self.download_file(download_url, zip_path)
|
||||
logger.info(f"unzip {zip_path}")
|
||||
with zipfile.ZipFile(zip_path, 'r') as zip_ref:
|
||||
zip_ref.extractall('temp/')
|
||||
# move temp/sd-concepts-library-main/sd-concepts-library to download_path
|
||||
logger.info(f"move temp/{download_name}-main/{download_name} to {download_path}")
|
||||
shutil.move(f"temp/{download_name}-main/{download_name}", download_path)
|
||||
logger.info(f"delete {zip_path}")
|
||||
os.remove(zip_path)
|
||||
logger.info(f"delete temp/{download_name}-main/")
|
||||
shutil.rmtree(f"temp/{download_name}-main")
|
||||
else:
|
||||
if not self.check_file_available(file_path):
|
||||
logger.init(f'{file_path}', status="Downloading")
|
||||
self.download_file(download_url, file_path)
|
||||
self.init()
|
||||
return True
|
||||
|
||||
def download_all_models(self):
|
||||
for model in self.get_filtered_model_names(download_all = True):
|
||||
if not self.check_model_available(model):
|
||||
logger.init(f"{model}", status="Downloading")
|
||||
self.download_model(model)
|
||||
else:
|
||||
logger.info(f"{model} is already downloaded.")
|
||||
return True
|
||||
|
||||
def download_all_dependencies(self):
|
||||
for dependency in self.dependencies:
|
||||
if not self.check_dependency_available(dependency):
|
||||
logger.init(f"{dependency}",status="Downloading")
|
||||
self.download_dependency(dependency)
|
||||
else:
|
||||
logger.info(f"{dependency} is already installed.")
|
||||
return True
|
||||
|
||||
def download_all(self):
|
||||
self.download_all_dependencies()
|
||||
self.download_all_models()
|
||||
return True
|
||||
|
||||
def check_all_available(self):
|
||||
for model in self.models:
|
||||
if not self.check_available(self.get_model_files(model)):
|
||||
return False
|
||||
for dependency in self.dependencies:
|
||||
if not self.check_available(self.get_dependency_files(dependency)):
|
||||
return False
|
||||
return True
|
||||
|
||||
def check_model_available(self, model_name):
|
||||
if model_name not in self.models:
|
||||
return False
|
||||
return self.check_available(self.get_model_files(model_name))
|
||||
|
||||
def check_dependency_available(self, dependency_name):
|
||||
if dependency_name not in self.dependencies:
|
||||
return False
|
||||
return self.check_available(self.get_dependency_files(dependency_name))
|
||||
|
||||
def check_all_available(self):
|
||||
for model in self.models:
|
||||
if not self.check_model_available(model):
|
||||
return False
|
||||
for dependency in self.dependencies:
|
||||
if not self.check_dependency_available(dependency):
|
||||
return False
|
||||
return True
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
@ -1,48 +0,0 @@
|
||||
# Class realesrgan
|
||||
# Inputs:
|
||||
# - model
|
||||
# - device
|
||||
# - output_dir
|
||||
# - output_ext
|
||||
# outupts:
|
||||
# - output_images
|
||||
import PIL
|
||||
from torchvision import transforms
|
||||
import numpy as np
|
||||
import os
|
||||
import cv2
|
||||
|
||||
from nataili.util.save_sample import save_sample
|
||||
|
||||
class realesrgan:
|
||||
def __init__(self, model, device, output_dir, output_ext='jpg'):
|
||||
self.model = model
|
||||
self.device = device
|
||||
self.output_dir = output_dir
|
||||
self.output_ext = output_ext
|
||||
self.output_images = []
|
||||
|
||||
def generate(self, input_image):
|
||||
# load image
|
||||
img = cv2.imread(input_image, cv2.IMREAD_UNCHANGED)
|
||||
if len(img.shape) == 3 and img.shape[2] == 4:
|
||||
img_mode = 'RGBA'
|
||||
else:
|
||||
img_mode = None
|
||||
# upscale
|
||||
output, _ = self.model.enhance(img)
|
||||
if img_mode == 'RGBA': # RGBA images should be saved in png format
|
||||
self.output_ext = 'png'
|
||||
|
||||
esrgan_sample = output[:,:,::-1]
|
||||
esrgan_image = PIL.Image.fromarray(esrgan_sample)
|
||||
# append model name to output image name
|
||||
filename = os.path.basename(input_image)
|
||||
filename = os.path.splitext(filename)[0]
|
||||
filename = f'{filename}_esrgan'
|
||||
filename_with_ext = f'{filename}.{self.output_ext}'
|
||||
output_image = os.path.join(self.output_dir, filename_with_ext)
|
||||
save_sample(esrgan_image, filename, self.output_dir, self.output_ext)
|
||||
self.output_images.append(output_image)
|
||||
return
|
||||
|
||||
@ -1,48 +0,0 @@
|
||||
# Class realesrgan
|
||||
# Inputs:
|
||||
# - model
|
||||
# - device
|
||||
# - output_dir
|
||||
# - output_ext
|
||||
# outupts:
|
||||
# - output_images
|
||||
import PIL
|
||||
from torchvision import transforms
|
||||
import numpy as np
|
||||
import os
|
||||
import cv2
|
||||
|
||||
from nataili.util.save_sample import save_sample
|
||||
|
||||
class realesrgan:
|
||||
def __init__(self, model, device, output_dir, output_ext='jpg'):
|
||||
self.model = model
|
||||
self.device = device
|
||||
self.output_dir = output_dir
|
||||
self.output_ext = output_ext
|
||||
self.output_images = []
|
||||
|
||||
def generate(self, input_image):
|
||||
# load image
|
||||
img = cv2.imread(input_image, cv2.IMREAD_UNCHANGED)
|
||||
if len(img.shape) == 3 and img.shape[2] == 4:
|
||||
img_mode = 'RGBA'
|
||||
else:
|
||||
img_mode = None
|
||||
# upscale
|
||||
output, _ = self.model.enhance(img)
|
||||
if img_mode == 'RGBA': # RGBA images should be saved in png format
|
||||
self.output_ext = 'png'
|
||||
|
||||
esrgan_sample = output[:,:,::-1]
|
||||
esrgan_image = PIL.Image.fromarray(esrgan_sample)
|
||||
# append model name to output image name
|
||||
filename = os.path.basename(input_image)
|
||||
filename = os.path.splitext(filename)[0]
|
||||
filename = f'{filename}_esrgan'
|
||||
filename_with_ext = f'{filename}.{self.output_ext}'
|
||||
output_image = os.path.join(self.output_dir, filename_with_ext)
|
||||
save_sample(esrgan_image, filename, self.output_dir, self.output_ext)
|
||||
self.output_images.append(output_image)
|
||||
return
|
||||
|
||||
@ -1 +0,0 @@
|
||||
from nataili.util.logger import logger,set_logger_verbosity, quiesce_logger, test_logger
|
||||
@ -1,16 +0,0 @@
|
||||
import gc
|
||||
|
||||
import torch
|
||||
import threading
|
||||
import pynvml
|
||||
import time
|
||||
|
||||
with torch.no_grad():
|
||||
def torch_gc():
|
||||
for _ in range(2):
|
||||
gc.collect()
|
||||
torch.cuda.empty_cache()
|
||||
torch.cuda.ipc_collect()
|
||||
torch.cuda.synchronize()
|
||||
torch.cuda.reset_peak_memory_stats()
|
||||
torch.cuda.reset_accumulated_memory_stats()
|
||||
@ -1,18 +0,0 @@
|
||||
def check_prompt_length(model, prompt, comments):
|
||||
"""this function tests if prompt is too long, and if so, adds a message to comments"""
|
||||
|
||||
tokenizer = model.cond_stage_model.tokenizer
|
||||
max_length = model.cond_stage_model.max_length
|
||||
|
||||
info = model.cond_stage_model.tokenizer([prompt], truncation=True, max_length=max_length,
|
||||
return_overflowing_tokens=True, padding="max_length", return_tensors="pt")
|
||||
ovf = info['overflowing_tokens'][0]
|
||||
overflowing_count = ovf.shape[0]
|
||||
if overflowing_count == 0:
|
||||
return
|
||||
|
||||
vocab = {v: k for k, v in tokenizer.get_vocab().items()}
|
||||
overflowing_words = [vocab.get(int(x), "") for x in ovf]
|
||||
overflowing_text = tokenizer.convert_tokens_to_string(''.join(overflowing_words))
|
||||
comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
|
||||
del tokenizer
|
||||
@ -1,22 +0,0 @@
|
||||
from pathlib import Path
|
||||
|
||||
def get_next_sequence_number(path, prefix=''):
|
||||
"""
|
||||
Determines and returns the next sequence number to use when saving an
|
||||
image in the specified directory.
|
||||
|
||||
If a prefix is given, only consider files whose names start with that
|
||||
prefix, and strip the prefix from filenames before extracting their
|
||||
sequence number.
|
||||
|
||||
The sequence starts at 0.
|
||||
"""
|
||||
result = -1
|
||||
for p in Path(path).iterdir():
|
||||
if p.name.endswith(('.png', '.jpg')) and p.name.startswith(prefix):
|
||||
tmp = p.name[len(prefix):]
|
||||
try:
|
||||
result = max(int(tmp.split('-')[0]), result)
|
||||
except ValueError:
|
||||
pass
|
||||
return result + 1
|
||||
@ -1,21 +0,0 @@
|
||||
import math
|
||||
|
||||
import PIL
|
||||
|
||||
|
||||
def image_grid(imgs, n_rows=None):
|
||||
if n_rows is not None:
|
||||
rows = n_rows
|
||||
else:
|
||||
rows = math.sqrt(len(imgs))
|
||||
rows = round(rows)
|
||||
|
||||
cols = math.ceil(len(imgs) / rows)
|
||||
|
||||
w, h = imgs[0].size
|
||||
grid = PIL.Image.new('RGB', size=(cols * w, rows * h), color='black')
|
||||
|
||||
for i, img in enumerate(imgs):
|
||||
grid.paste(img, box=(i % cols * w, i // cols * h))
|
||||
|
||||
return grid
|
||||
@ -1,40 +0,0 @@
|
||||
import os
|
||||
|
||||
import torch
|
||||
|
||||
|
||||
def load_learned_embed_in_clip(learned_embeds_path, text_encoder, tokenizer, token=None):
|
||||
loaded_learned_embeds = torch.load(learned_embeds_path, map_location="cpu")
|
||||
# separate token and the embeds
|
||||
if learned_embeds_path.endswith('.pt'):
|
||||
# old format
|
||||
# token = * so replace with file directory name when converting
|
||||
trained_token = os.path.basename(learned_embeds_path)
|
||||
params_dict = {
|
||||
trained_token: torch.tensor(list(loaded_learned_embeds['string_to_param'].items())[0][1])
|
||||
}
|
||||
learned_embeds_path = os.path.splitext(learned_embeds_path)[0] + '.bin'
|
||||
torch.save(params_dict, learned_embeds_path)
|
||||
loaded_learned_embeds = torch.load(learned_embeds_path, map_location="cpu")
|
||||
trained_token = list(loaded_learned_embeds.keys())[0]
|
||||
embeds = loaded_learned_embeds[trained_token]
|
||||
elif learned_embeds_path.endswith('.bin'):
|
||||
trained_token = list(loaded_learned_embeds.keys())[0]
|
||||
embeds = loaded_learned_embeds[trained_token]
|
||||
|
||||
embeds = loaded_learned_embeds[trained_token]
|
||||
# cast to dtype of text_encoder
|
||||
dtype = text_encoder.get_input_embeddings().weight.dtype
|
||||
embeds.to(dtype)
|
||||
|
||||
# add the token in tokenizer
|
||||
token = token if token is not None else trained_token
|
||||
num_added_tokens = tokenizer.add_tokens(token)
|
||||
|
||||
# resize the token embeddings
|
||||
text_encoder.resize_token_embeddings(len(tokenizer))
|
||||
|
||||
# get the id for the token and assign the embeds
|
||||
token_id = tokenizer.convert_tokens_to_ids(token)
|
||||
text_encoder.get_input_embeddings().weight.data[token_id] = embeds
|
||||
return token
|
||||
@ -1,102 +0,0 @@
|
||||
import sys
|
||||
from functools import partialmethod
|
||||
from loguru import logger
|
||||
|
||||
STDOUT_LEVELS = ["GENERATION", "PROMPT"]
|
||||
INIT_LEVELS = ["INIT", "INIT_OK", "INIT_WARN", "INIT_ERR"]
|
||||
MESSAGE_LEVELS = ["MESSAGE"]
|
||||
# By default we're at error level or higher
|
||||
verbosity = 20
|
||||
quiet = 0
|
||||
|
||||
def set_logger_verbosity(count):
|
||||
global verbosity
|
||||
# The count comes reversed. So count = 0 means minimum verbosity
|
||||
# While count 5 means maximum verbosity
|
||||
# So the more count we have, the lowe we drop the versbosity maximum
|
||||
verbosity = 20 - (count * 10)
|
||||
|
||||
def quiesce_logger(count):
|
||||
global quiet
|
||||
# The bigger the count, the more silent we want our logger
|
||||
quiet = count * 10
|
||||
|
||||
def is_stdout_log(record):
|
||||
if record["level"].name not in STDOUT_LEVELS:
|
||||
return(False)
|
||||
if record["level"].no < verbosity + quiet:
|
||||
return(False)
|
||||
return(True)
|
||||
|
||||
def is_init_log(record):
|
||||
if record["level"].name not in INIT_LEVELS:
|
||||
return(False)
|
||||
if record["level"].no < verbosity + quiet:
|
||||
return(False)
|
||||
return(True)
|
||||
|
||||
def is_msg_log(record):
|
||||
if record["level"].name not in MESSAGE_LEVELS:
|
||||
return(False)
|
||||
if record["level"].no < verbosity + quiet:
|
||||
return(False)
|
||||
return(True)
|
||||
|
||||
def is_stderr_log(record):
|
||||
if record["level"].name in STDOUT_LEVELS + INIT_LEVELS + MESSAGE_LEVELS:
|
||||
return(False)
|
||||
if record["level"].no < verbosity + quiet:
|
||||
return(False)
|
||||
return(True)
|
||||
|
||||
def test_logger():
|
||||
logger.generation("This is a generation message\nIt is typically multiline\nThee Lines".encode("unicode_escape").decode("utf-8"))
|
||||
logger.prompt("This is a prompt message")
|
||||
logger.debug("Debug Message")
|
||||
logger.info("Info Message")
|
||||
logger.warning("Info Warning")
|
||||
logger.error("Error Message")
|
||||
logger.critical("Critical Message")
|
||||
logger.init("This is an init message", status="Starting")
|
||||
logger.init_ok("This is an init message", status="OK")
|
||||
logger.init_warn("This is an init message", status="Warning")
|
||||
logger.init_err("This is an init message", status="Error")
|
||||
logger.message("This is user message")
|
||||
sys.exit()
|
||||
|
||||
|
||||
logfmt = "<level>{level: <10}</level> | <green>{time:YYYY-MM-DD HH:mm:ss}</green> | <green>{name}</green>:<green>{function}</green>:<green>{line}</green> - <level>{message}</level>"
|
||||
genfmt = "<level>{level: <10}</level> @ <green>{time:YYYY-MM-DD HH:mm:ss}</green> | <level>{message}</level>"
|
||||
initfmt = "<magenta>INIT </magenta> | <level>{extra[status]: <11}</level> | <magenta>{message}</magenta>"
|
||||
msgfmt = "<level>{level: <10}</level> | <level>{message}</level>"
|
||||
|
||||
try:
|
||||
logger.level("GENERATION", no=24, color="<cyan>")
|
||||
logger.level("PROMPT", no=23, color="<yellow>")
|
||||
logger.level("INIT", no=31, color="<white>")
|
||||
logger.level("INIT_OK", no=31, color="<green>")
|
||||
logger.level("INIT_WARN", no=31, color="<yellow>")
|
||||
logger.level("INIT_ERR", no=31, color="<red>")
|
||||
# Messages contain important information without which this application might not be able to be used
|
||||
# As such, they have the highest priority
|
||||
logger.level("MESSAGE", no=61, color="<green>")
|
||||
except TypeError:
|
||||
pass
|
||||
|
||||
logger.__class__.generation = partialmethod(logger.__class__.log, "GENERATION")
|
||||
logger.__class__.prompt = partialmethod(logger.__class__.log, "PROMPT")
|
||||
logger.__class__.init = partialmethod(logger.__class__.log, "INIT")
|
||||
logger.__class__.init_ok = partialmethod(logger.__class__.log, "INIT_OK")
|
||||
logger.__class__.init_warn = partialmethod(logger.__class__.log, "INIT_WARN")
|
||||
logger.__class__.init_err = partialmethod(logger.__class__.log, "INIT_ERR")
|
||||
logger.__class__.message = partialmethod(logger.__class__.log, "MESSAGE")
|
||||
|
||||
config = {
|
||||
"handlers": [
|
||||
{"sink": sys.stderr, "format": logfmt, "colorize":True, "filter": is_stderr_log},
|
||||
{"sink": sys.stdout, "format": genfmt, "level": "PROMPT", "colorize":True, "filter": is_stdout_log},
|
||||
{"sink": sys.stdout, "format": initfmt, "level": "INIT", "colorize":True, "filter": is_init_log},
|
||||
{"sink": sys.stdout, "format": msgfmt, "level": "MESSAGE", "colorize":True, "filter": is_msg_log}
|
||||
],
|
||||
}
|
||||
logger.configure(**config)
|
||||
@ -1,20 +0,0 @@
|
||||
import os
|
||||
|
||||
def save_sample(image, filename, sample_path, extension='png', jpg_quality=95, webp_quality=95, webp_lossless=True, png_compression=9):
|
||||
path = os.path.join(sample_path, filename + '.' + extension)
|
||||
if os.path.exists(path):
|
||||
return False
|
||||
if not os.path.exists(sample_path):
|
||||
os.makedirs(sample_path)
|
||||
if extension == 'png':
|
||||
image.save(path, format='PNG', compress_level=png_compression)
|
||||
elif extension == 'jpg':
|
||||
image.save(path, quality=jpg_quality, optimize=True)
|
||||
elif extension == 'webp':
|
||||
image.save(path, quality=webp_quality, lossless=webp_lossless)
|
||||
else:
|
||||
return False
|
||||
if os.path.exists(path):
|
||||
return True
|
||||
else:
|
||||
return False
|
||||
@ -1,22 +0,0 @@
|
||||
import random
|
||||
|
||||
def seed_to_int(s):
|
||||
if type(s) is int:
|
||||
return s
|
||||
if s is None or s == '':
|
||||
return random.randint(0, 2**32 - 1)
|
||||
|
||||
if type(s) is list:
|
||||
seed_list = []
|
||||
for seed in s:
|
||||
if seed is None or seed == '':
|
||||
seed_list.append(random.randint(0, 2**32 - 1))
|
||||
else:
|
||||
seed_list = s
|
||||
|
||||
return seed_list
|
||||
|
||||
n = abs(int(s) if s.isdigit() else random.Random(s).randint(0, 2**32 - 1))
|
||||
while n >= 2**32:
|
||||
n = n >> 32
|
||||
return n
|
||||
34
scripts/post_processing.py
Normal file
34
scripts/post_processing.py
Normal file
@ -0,0 +1,34 @@
|
||||
# This file is part of sygil-webui (https://github.com/Sygil-Dev/sandbox-webui/).
|
||||
|
||||
# Copyright 2022 Sygil-Dev team.
|
||||
# This program is free software: you can redistribute it and/or modify
|
||||
# it under the terms of the GNU Affero General Public License as published by
|
||||
# the Free Software Foundation, either version 3 of the License, or
|
||||
# (at your option) any later version.
|
||||
|
||||
# This program is distributed in the hope that it will be useful,
|
||||
# but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
# GNU Affero General Public License for more details.
|
||||
|
||||
# You should have received a copy of the GNU Affero General Public License
|
||||
# along with this program. If not, see <http://www.gnu.org/licenses/>.
|
||||
# base webui import and utils.
|
||||
#from sd_utils import *
|
||||
from sd_utils import *
|
||||
# streamlit imports
|
||||
|
||||
#streamlit components section
|
||||
|
||||
#other imports
|
||||
import os, time, requests
|
||||
import sys
|
||||
|
||||
# Temp imports
|
||||
|
||||
# end of imports
|
||||
#---------------------------------------------------------------------------------------------------------------
|
||||
|
||||
|
||||
def layout():
|
||||
st.info("Under Construction. :construction_worker:")
|
||||
@ -14,15 +14,13 @@
|
||||
# You should have received a copy of the GNU Affero General Public License
|
||||
# along with this program. If not, see <http://www.gnu.org/licenses/>.
|
||||
# base webui import and utils.
|
||||
import collections.abc
|
||||
#from webui_streamlit import st
|
||||
import gfpgan
|
||||
import hydralit as st
|
||||
|
||||
|
||||
# streamlit imports
|
||||
from streamlit import StopException, StreamlitAPIException
|
||||
from streamlit.runtime.scriptrunner import script_run_context
|
||||
#from streamlit.runtime.scriptrunner import script_run_context
|
||||
|
||||
#streamlit components section
|
||||
from streamlit_server_state import server_state, server_state_lock
|
||||
@ -35,7 +33,7 @@ import streamlit_nested_layout
|
||||
import warnings
|
||||
import json
|
||||
|
||||
import base64
|
||||
import base64, cv2
|
||||
import os, sys, re, random, datetime, time, math, glob, toml
|
||||
import gc
|
||||
from PIL import Image, ImageFont, ImageDraw, ImageFilter
|
||||
@ -68,15 +66,25 @@ import piexif.helper
|
||||
from tqdm import trange
|
||||
from ldm.models.diffusion.ddim import DDIMSampler
|
||||
from ldm.util import ismap
|
||||
from abc import ABC, abstractmethod
|
||||
#from abc import ABC, abstractmethod
|
||||
from typing import Dict, Union
|
||||
from io import BytesIO
|
||||
from packaging import version
|
||||
from uuid import uuid4
|
||||
from pathlib import Path
|
||||
from huggingface_hub import hf_hub_download
|
||||
|
||||
#import librosa
|
||||
from logger import logger, set_logger_verbosity, quiesce_logger
|
||||
#from loguru import logger
|
||||
|
||||
try:
|
||||
from realesrgan import RealESRGANer
|
||||
from basicsr.archs.rrdbnet_arch import RRDBNet
|
||||
except ImportError as e:
|
||||
logger.error("You tried to import realesrgan without having it installed properly. To install Real-ESRGAN, run:\n\n"
|
||||
"pip install realesrgan")
|
||||
|
||||
# Temp imports
|
||||
#from basicsr.utils.registry import ARCH_REGISTRY
|
||||
|
||||
@ -84,14 +92,6 @@ from logger import logger, set_logger_verbosity, quiesce_logger
|
||||
# end of imports
|
||||
#---------------------------------------------------------------------------------------------------------------
|
||||
|
||||
# we make a log file where we store the logs
|
||||
logger.add("logs/log_{time:MM-DD-YYYY!UTC}.log", rotation="8 MB", compression="zip", level='INFO') # Once the file is too old, it's rotated
|
||||
logger.add(sys.stderr, diagnose=True)
|
||||
logger.add(sys.stdout)
|
||||
logger.enable("")
|
||||
|
||||
#
|
||||
|
||||
try:
|
||||
# this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start.
|
||||
from transformers import logging
|
||||
@ -112,6 +112,8 @@ mimetypes.add_type('application/javascript', '.js')
|
||||
opt_C = 4
|
||||
opt_f = 8
|
||||
|
||||
# The model manager loads and unloads the SD models and has features to download them or find their location
|
||||
#model_manager = ModelManager()
|
||||
|
||||
def load_configs():
|
||||
if not "defaults" in st.session_state:
|
||||
@ -269,6 +271,33 @@ def make_grid(n_items=5, n_cols=5):
|
||||
|
||||
return cols
|
||||
|
||||
|
||||
def merge(file1, file2, out, weight):
|
||||
alpha = (weight)/100
|
||||
if not(file1.endswith(".ckpt")):
|
||||
file1 += ".ckpt"
|
||||
if not(file2.endswith(".ckpt")):
|
||||
file2 += ".ckpt"
|
||||
if not(out.endswith(".ckpt")):
|
||||
out += ".ckpt"
|
||||
#Load Models
|
||||
model_0 = torch.load(file1)
|
||||
model_1 = torch.load(file2)
|
||||
theta_0 = model_0['state_dict']
|
||||
theta_1 = model_1['state_dict']
|
||||
|
||||
for key in theta_0.keys():
|
||||
if 'model' in key and key in theta_1:
|
||||
theta_0[key] = (alpha) * theta_0[key] + (1-alpha) * theta_1[key]
|
||||
|
||||
logger.info("RUNNING...\n(STAGE 2)")
|
||||
|
||||
for key in theta_1.keys():
|
||||
if 'model' in key and key not in theta_0:
|
||||
theta_0[key] = theta_1[key]
|
||||
torch.save(model_0, out)
|
||||
|
||||
|
||||
def human_readable_size(size, decimal_places=3):
|
||||
"""Return a human readable size from bytes."""
|
||||
for unit in ['B','KB','MB','GB','TB']:
|
||||
@ -282,6 +311,8 @@ def load_models(use_LDSR = False, LDSR_model='model', use_GFPGAN=False, GFPGAN_m
|
||||
CustomModel_available=False, custom_model="Stable Diffusion v1.5"):
|
||||
"""Load the different models. We also reuse the models that are already in memory to speed things up instead of loading them again. """
|
||||
|
||||
#model_manager.init()
|
||||
|
||||
logger.info("Loading models.")
|
||||
|
||||
if "progress_bar_text" in st.session_state:
|
||||
@ -431,7 +462,6 @@ def load_models(use_LDSR = False, LDSR_model='model', use_GFPGAN=False, GFPGAN_m
|
||||
try:
|
||||
server_state["model"].args.use_multiprocessing_for_evaluation = False
|
||||
except AttributeError as e:
|
||||
logger.error(e)
|
||||
pass
|
||||
|
||||
if st.session_state.defaults.general.enable_attention_slicing:
|
||||
@ -1350,6 +1380,77 @@ def load_RealESRGAN(model_name: str):
|
||||
|
||||
return server_state['RealESRGAN']
|
||||
|
||||
#
|
||||
class RealESRGANModel(nn.Module):
|
||||
def __init__(self, model_path, tile=0, tile_pad=10, pre_pad=0, fp32=False):
|
||||
super().__init__()
|
||||
try:
|
||||
from basicsr.archs.rrdbnet_arch import RRDBNet
|
||||
from realesrgan import RealESRGANer
|
||||
except ImportError as e:
|
||||
logger.error(
|
||||
"You tried to import realesrgan without having it installed properly. To install Real-ESRGAN, run:\n\n"
|
||||
"pip install realesrgan"
|
||||
)
|
||||
|
||||
model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
|
||||
self.upsampler = RealESRGANer(
|
||||
scale=4, model_path=model_path, model=model, tile=tile, tile_pad=tile_pad, pre_pad=pre_pad, half=not fp32
|
||||
)
|
||||
|
||||
def forward(self, image, outscale=4, convert_to_pil=True):
|
||||
"""Upsample an image array or path.
|
||||
Args:
|
||||
image (Union[np.ndarray, str]): Either a np array or an image path. np array is assumed to be in RGB format,
|
||||
and we convert it to BGR.
|
||||
outscale (int, optional): Amount to upscale the image. Defaults to 4.
|
||||
convert_to_pil (bool, optional): If True, return PIL image. Otherwise, return numpy array (BGR). Defaults to True.
|
||||
Returns:
|
||||
Union[np.ndarray, PIL.Image.Image]: An upsampled version of the input image.
|
||||
"""
|
||||
if isinstance(image, (str, Path)):
|
||||
img = cv2.imread(image, cv2.IMREAD_UNCHANGED)
|
||||
else:
|
||||
img = image
|
||||
img = (img * 255).round().astype("uint8")
|
||||
img = img[:, :, ::-1]
|
||||
|
||||
image, _ = self.upsampler.enhance(img, outscale=outscale)
|
||||
|
||||
if convert_to_pil:
|
||||
image = Image.fromarray(image[:, :, ::-1])
|
||||
|
||||
return image
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, model_name_or_path="nateraw/real-esrgan"):
|
||||
"""Initialize a pretrained Real-ESRGAN upsampler.
|
||||
Args:
|
||||
model_name_or_path (str, optional): The Hugging Face repo ID or path to local model. Defaults to 'nateraw/real-esrgan'.
|
||||
Returns:
|
||||
PipelineRealESRGAN: An instance of `PipelineRealESRGAN` instantiated from pretrained model.
|
||||
"""
|
||||
# reuploaded form official ones mentioned here:
|
||||
# https://github.com/xinntao/Real-ESRGAN
|
||||
if Path(model_name_or_path).exists():
|
||||
file = model_name_or_path
|
||||
else:
|
||||
file = hf_hub_download(model_name_or_path, "RealESRGAN_x4plus.pth")
|
||||
return cls(file)
|
||||
|
||||
def upsample_imagefolder(self, in_dir, out_dir, suffix="out", outfile_ext=".png"):
|
||||
in_dir, out_dir = Path(in_dir), Path(out_dir)
|
||||
if not in_dir.exists():
|
||||
raise FileNotFoundError(f"Provided input directory {in_dir} does not exist")
|
||||
|
||||
out_dir.mkdir(exist_ok=True, parents=True)
|
||||
|
||||
image_paths = [x for x in in_dir.glob("*") if x.suffix.lower() in [".png", ".jpg", ".jpeg"]]
|
||||
for image in image_paths:
|
||||
im = self(str(image))
|
||||
out_filepath = out_dir / (image.stem + suffix + outfile_ext)
|
||||
im.save(out_filepath)
|
||||
|
||||
#
|
||||
@retry(tries=5)
|
||||
def load_LDSR(model_name="model", config="project", checking=False):
|
||||
@ -1744,6 +1845,9 @@ def seed_to_int(s):
|
||||
if s is None or s == '':
|
||||
return random.randint(0, 2**32 - 1)
|
||||
|
||||
if ',' in s:
|
||||
s = s.split(',')
|
||||
|
||||
if type(s) is list:
|
||||
seed_list = []
|
||||
for seed in s:
|
||||
@ -1906,6 +2010,8 @@ def GFPGAN_available():
|
||||
st.session_state["GFPGAN_available"] = True
|
||||
else:
|
||||
st.session_state["GFPGAN_available"] = False
|
||||
st.session_state["use_GFPGAN"] = False
|
||||
st.session_state["GFPGAN_model"] = "GFPGANv1.4"
|
||||
|
||||
#
|
||||
def RealESRGAN_available():
|
||||
@ -1924,6 +2030,8 @@ def RealESRGAN_available():
|
||||
st.session_state["RealESRGAN_available"] = True
|
||||
else:
|
||||
st.session_state["RealESRGAN_available"] = False
|
||||
st.session_state["use_RealESRGAN"] = False
|
||||
st.session_state["RealESRGAN_model"] = "RealESRGAN_x4plus"
|
||||
#
|
||||
def LDSR_available():
|
||||
#with server_state_lock["RealESRGAN_models"]:
|
||||
@ -1944,6 +2052,8 @@ def LDSR_available():
|
||||
st.session_state["LDSR_available"] = True
|
||||
else:
|
||||
st.session_state["LDSR_available"] = False
|
||||
st.session_state["use_LDSR"] = False
|
||||
st.session_state["LDSR_model"] = "model"
|
||||
|
||||
|
||||
|
||||
@ -1955,6 +2065,7 @@ def save_sample(image, sample_path_i, filename, jpg_sample, prompts, seeds, widt
|
||||
|
||||
filename_i = os.path.join(sample_path_i, filename)
|
||||
|
||||
if "defaults" in st.session_state:
|
||||
if st.session_state['defaults'].general.save_metadata or write_info_files:
|
||||
# toggles differ for txt2img vs. img2img:
|
||||
offset = 0 if init_img is None else 2
|
||||
@ -2563,7 +2674,7 @@ def process_images(
|
||||
#output_images.append(image)
|
||||
#if simple_templating:
|
||||
#grid_captions.append( captions[i] )
|
||||
|
||||
if "defaults" in st.session_state:
|
||||
if st.session_state['defaults'].general.optimized:
|
||||
mem = torch.cuda.memory_allocated()/1e6
|
||||
server_state["modelFS"].to("cpu")
|
||||
|
||||
@ -10,7 +10,6 @@ import time
|
||||
import json
|
||||
|
||||
import torch
|
||||
from diffusers import ModelMixin
|
||||
from diffusers.configuration_utils import FrozenDict
|
||||
from diffusers.models import AutoencoderKL, UNet2DConditionModel
|
||||
from diffusers.pipeline_utils import DiffusionPipeline
|
||||
@ -22,59 +21,39 @@ from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
|
||||
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
|
||||
from torch import nn
|
||||
|
||||
from .upsampling import RealESRGANModel
|
||||
|
||||
from sd_utils import RealESRGANModel
|
||||
|
||||
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
||||
|
||||
|
||||
def get_spec_norm(wav, sr, n_mels=512, hop_length=704):
|
||||
"""Obtain maximum value for each time-frame in Mel Spectrogram,
|
||||
and normalize between 0 and 1
|
||||
def get_timesteps_arr(audio_filepath, offset, duration, fps=30, margin=1.0, smooth=0.0):
|
||||
y, sr = librosa.load(audio_filepath, offset=offset, duration=duration)
|
||||
|
||||
Borrowed from lucid sonic dreams repo. In there, they programatically determine hop length
|
||||
but I really didn't understand what was going on so I removed it and hard coded the output.
|
||||
"""
|
||||
# librosa.stft hardcoded defaults...
|
||||
# n_fft defaults to 2048
|
||||
# hop length is win_length // 4
|
||||
# win_length defaults to n_fft
|
||||
D = librosa.stft(y, n_fft=2048, hop_length=2048 // 4, win_length=2048)
|
||||
|
||||
# Generate Mel Spectrogram
|
||||
spec_raw = librosa.feature.melspectrogram(y=wav, sr=sr, n_mels=n_mels, hop_length=hop_length)
|
||||
# Extract percussive elements
|
||||
D_harmonic, D_percussive = librosa.decompose.hpss(D, margin=margin)
|
||||
y_percussive = librosa.istft(D_percussive, length=len(y))
|
||||
|
||||
# Obtain maximum value per time-frame
|
||||
# Get normalized melspectrogram
|
||||
spec_raw = librosa.feature.melspectrogram(y=y_percussive, sr=sr)
|
||||
spec_max = np.amax(spec_raw, axis=0)
|
||||
|
||||
# Normalize all values between 0 and 1
|
||||
spec_norm = (spec_max - np.min(spec_max)) / np.ptp(spec_max)
|
||||
|
||||
return spec_norm
|
||||
# Resize cumsum of spec norm to our desired number of interpolation frames
|
||||
x_norm = np.linspace(0, spec_norm.shape[-1], spec_norm.shape[-1])
|
||||
y_norm = np.cumsum(spec_norm)
|
||||
y_norm /= y_norm[-1]
|
||||
x_resize = np.linspace(0, y_norm.shape[-1], int(duration*fps))
|
||||
|
||||
T = np.interp(x_resize, x_norm, y_norm)
|
||||
|
||||
def get_timesteps_arr(audio_filepath, offset, duration, fps=30, margin=(1.0, 5.0)):
|
||||
"""Get the array that will be used to determine how much to interpolate between images.
|
||||
|
||||
Normally, this is just a linspace between 0 and 1 for the number of frames to generate. In this case,
|
||||
we want to use the amplitude of the audio to determine how much to interpolate between images.
|
||||
|
||||
So, here we:
|
||||
1. Load the audio file
|
||||
2. Split the audio into harmonic and percussive components
|
||||
3. Get the normalized amplitude of the percussive component, resized to the number of frames
|
||||
4. Get the cumulative sum of the amplitude array
|
||||
5. Normalize the cumulative sum between 0 and 1
|
||||
6. Return the array
|
||||
|
||||
I honestly have no clue what I'm doing here. Suggestions welcome.
|
||||
"""
|
||||
y, sr = librosa.load(audio_filepath, offset=offset, duration=duration)
|
||||
wav_harmonic, wav_percussive = librosa.effects.hpss(y, margin=margin)
|
||||
|
||||
# Apparently n_mels is supposed to be input shape but I don't think it matters here?
|
||||
frame_duration = int(sr / fps)
|
||||
wav_norm = get_spec_norm(wav_percussive, sr, n_mels=512, hop_length=frame_duration)
|
||||
amplitude_arr = np.resize(wav_norm, int(duration * fps))
|
||||
T = np.cumsum(amplitude_arr)
|
||||
T /= T[-1]
|
||||
T[0] = 0.0
|
||||
return T
|
||||
# Apply smoothing
|
||||
return T * (1 - smooth) + np.linspace(0.0, 1.0, T.shape[0]) * smooth
|
||||
|
||||
|
||||
def slerp(t, v0, v1, DOT_THRESHOLD=0.9995):
|
||||
@ -130,7 +109,6 @@ def make_video_pyav(
|
||||
frame = pil_to_tensor(Image.open(img)).unsqueeze(0)
|
||||
frames = frame if frames is None else torch.cat([frames, frame])
|
||||
else:
|
||||
|
||||
frames = frames_or_frame_dir
|
||||
|
||||
# TCHW -> THWC
|
||||
@ -208,6 +186,16 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
new_config["steps_offset"] = 1
|
||||
scheduler._internal_dict = FrozenDict(new_config)
|
||||
|
||||
if safety_checker is None:
|
||||
logger.warn(
|
||||
f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
|
||||
" that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
|
||||
" results in services or applications open to the public. Both the diffusers team and Hugging Face"
|
||||
" strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
|
||||
" it only for use-cases that involve analyzing network behavior or auditing its results. For more"
|
||||
" information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
|
||||
)
|
||||
|
||||
self.register_modules(
|
||||
vae=vae,
|
||||
text_encoder=text_encoder,
|
||||
@ -251,6 +239,8 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
width: int = 512,
|
||||
num_inference_steps: int = 50,
|
||||
guidance_scale: float = 7.5,
|
||||
negative_prompt: Optional[Union[str, List[str]]] = None,
|
||||
num_images_per_prompt: Optional[int] = 1,
|
||||
eta: float = 0.0,
|
||||
generator: Optional[torch.Generator] = None,
|
||||
latents: Optional[torch.FloatTensor] = None,
|
||||
@ -259,12 +249,13 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
||||
callback_steps: Optional[int] = 1,
|
||||
text_embeddings: Optional[torch.FloatTensor] = None,
|
||||
**kwargs,
|
||||
):
|
||||
r"""
|
||||
Function invoked when calling the pipeline for generation.
|
||||
Args:
|
||||
prompt (`str` or `List[str]`):
|
||||
The prompt or prompts to guide the image generation.
|
||||
prompt (`str` or `List[str]`, *optional*, defaults to `None`):
|
||||
The prompt or prompts to guide the image generation. If not provided, `text_embeddings` is required.
|
||||
height (`int`, *optional*, defaults to 512):
|
||||
The height in pixels of the generated image.
|
||||
width (`int`, *optional*, defaults to 512):
|
||||
@ -278,6 +269,11 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
||||
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
||||
usually at the expense of lower image quality.
|
||||
negative_prompt (`str` or `List[str]`, *optional*):
|
||||
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
||||
if `guidance_scale` is less than `1`).
|
||||
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
||||
The number of images to generate per prompt.
|
||||
eta (`float`, *optional*, defaults to 0.0):
|
||||
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
|
||||
[`schedulers.DDIMScheduler`], will be ignored for others.
|
||||
@ -300,8 +296,10 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
callback_steps (`int`, *optional*, defaults to 1):
|
||||
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
||||
called at every step.
|
||||
text_embeddings(`torch.FloatTensor`, *optional*):
|
||||
Pre-generated text embeddings.
|
||||
text_embeddings (`torch.FloatTensor`, *optional*, defaults to `None`):
|
||||
Pre-generated text embeddings to be used as inputs for image generation. Can be used in place of
|
||||
`prompt` to avoid re-computing the embeddings. If not provided, the embeddings will be generated from
|
||||
the supplied `prompt`.
|
||||
Returns:
|
||||
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
||||
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
|
||||
@ -340,7 +338,7 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
|
||||
if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
|
||||
removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
|
||||
logger.warning(
|
||||
print(
|
||||
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
||||
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
|
||||
)
|
||||
@ -349,21 +347,51 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
else:
|
||||
batch_size = text_embeddings.shape[0]
|
||||
|
||||
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
||||
bs_embed, seq_len, _ = text_embeddings.shape
|
||||
text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
|
||||
text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
||||
|
||||
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
||||
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
||||
# corresponds to doing no classifier free guidance.
|
||||
do_classifier_free_guidance = guidance_scale > 1.0
|
||||
# get unconditional embeddings for classifier free guidance
|
||||
if do_classifier_free_guidance:
|
||||
# HACK - Not setting text_input_ids here when walking, so hard coding to max length of tokenizer
|
||||
# TODO - Determine if this is OK to do
|
||||
# max_length = text_input_ids.shape[-1]
|
||||
uncond_tokens: List[str]
|
||||
if negative_prompt is None:
|
||||
uncond_tokens = [""]
|
||||
elif type(prompt) is not type(negative_prompt):
|
||||
raise TypeError(
|
||||
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
||||
f" {type(prompt)}."
|
||||
)
|
||||
elif isinstance(negative_prompt, str):
|
||||
uncond_tokens = [negative_prompt]
|
||||
elif batch_size != len(negative_prompt):
|
||||
raise ValueError(
|
||||
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
||||
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
||||
" the batch size of `prompt`."
|
||||
)
|
||||
else:
|
||||
uncond_tokens = negative_prompt
|
||||
|
||||
max_length = self.tokenizer.model_max_length
|
||||
uncond_input = self.tokenizer(
|
||||
[""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
|
||||
uncond_tokens,
|
||||
padding="max_length",
|
||||
max_length=max_length,
|
||||
truncation=True,
|
||||
return_tensors="pt",
|
||||
)
|
||||
uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
|
||||
|
||||
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
|
||||
seq_len = uncond_embeddings.shape[1]
|
||||
uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1)
|
||||
uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
|
||||
|
||||
# For classifier free guidance, we need to do two forward passes.
|
||||
# Here we concatenate the unconditional and text embeddings into a single batch
|
||||
# to avoid doing two forward passes
|
||||
@ -374,19 +402,20 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
# Unlike in other pipelines, latents need to be generated in the target device
|
||||
# for 1-to-1 results reproducibility with the CompVis implementation.
|
||||
# However this currently doesn't work in `mps`.
|
||||
latents_device = "cpu" if self.device.type == "mps" else self.device
|
||||
latents_shape = (batch_size, self.unet.in_channels, height // 8, width // 8)
|
||||
latents_shape = (batch_size * num_images_per_prompt, self.unet.in_channels, height // 8, width // 8)
|
||||
latents_dtype = text_embeddings.dtype
|
||||
if latents is None:
|
||||
latents = torch.randn(
|
||||
latents_shape,
|
||||
generator=generator,
|
||||
device=latents_device,
|
||||
dtype=text_embeddings.dtype,
|
||||
if self.device.type == "mps":
|
||||
# randn does not exist on mps
|
||||
latents = torch.randn(latents_shape, generator=generator, device="cpu", dtype=latents_dtype).to(
|
||||
self.device
|
||||
)
|
||||
else:
|
||||
latents = torch.randn(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
|
||||
else:
|
||||
if latents.shape != latents_shape:
|
||||
raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
|
||||
latents = latents.to(latents_device)
|
||||
latents = latents.to(self.device)
|
||||
|
||||
# set timesteps
|
||||
self.scheduler.set_timesteps(num_inference_steps)
|
||||
@ -431,12 +460,19 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
image = self.vae.decode(latents).sample
|
||||
|
||||
image = (image / 2 + 0.5).clamp(0, 1)
|
||||
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
||||
|
||||
safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(self.device)
|
||||
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
|
||||
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
||||
|
||||
if self.safety_checker is not None:
|
||||
safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
|
||||
self.device
|
||||
)
|
||||
image, has_nsfw_concept = self.safety_checker(
|
||||
images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype)
|
||||
)
|
||||
else:
|
||||
has_nsfw_concept = None
|
||||
|
||||
if output_type == "pil":
|
||||
image = self.numpy_to_pil(image)
|
||||
@ -449,16 +485,9 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
def generate_inputs(self, prompt_a, prompt_b, seed_a, seed_b, noise_shape, T, batch_size):
|
||||
embeds_a = self.embed_text(prompt_a)
|
||||
embeds_b = self.embed_text(prompt_b)
|
||||
latents_a = torch.randn(
|
||||
noise_shape,
|
||||
device=self.device,
|
||||
generator=torch.Generator(device=self.device).manual_seed(seed_a),
|
||||
)
|
||||
latents_b = torch.randn(
|
||||
noise_shape,
|
||||
device=self.device,
|
||||
generator=torch.Generator(device=self.device).manual_seed(seed_b),
|
||||
)
|
||||
|
||||
latents_a = self.init_noise(seed_a, noise_shape)
|
||||
latents_b = self.init_noise(seed_b, noise_shape)
|
||||
|
||||
batch_idx = 0
|
||||
embeds_batch, noise_batch = None, None
|
||||
@ -477,7 +506,7 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
torch.cuda.empty_cache()
|
||||
embeds_batch, noise_batch = None, None
|
||||
|
||||
def generate_interpolation_clip(
|
||||
def make_clip_frames(
|
||||
self,
|
||||
prompt_a: str,
|
||||
prompt_b: str,
|
||||
@ -530,7 +559,7 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
eta=eta,
|
||||
num_inference_steps=num_inference_steps,
|
||||
output_type="pil" if not upsample else "numpy",
|
||||
)["sample"]
|
||||
)["images"]
|
||||
|
||||
for image in outputs:
|
||||
frame_filepath = save_path / (f"frame%06d{image_file_ext}" % frame_index)
|
||||
@ -557,6 +586,8 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
resume: Optional[bool] = False,
|
||||
audio_filepath: str = None,
|
||||
audio_start_sec: Optional[Union[int, float]] = None,
|
||||
margin: Optional[float] = 1.0,
|
||||
smooth: Optional[float] = 0.0,
|
||||
):
|
||||
"""Generate a video from a sequence of prompts and seeds. Optionally, add audio to the
|
||||
video to interpolate to the intensity of the audio.
|
||||
@ -603,13 +634,17 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
Optional path to an audio file to influence the interpolation rate.
|
||||
audio_start_sec (Optional[Union[int, float]], *optional*, defaults to 0):
|
||||
Global start time of the provided audio_filepath.
|
||||
margin (Optional[float], *optional*, defaults to 1.0):
|
||||
Margin from librosa hpss to use for audio interpolation.
|
||||
smooth (Optional[float], *optional*, defaults to 0.0):
|
||||
Smoothness of the audio interpolation. 1.0 means linear interpolation.
|
||||
|
||||
This function will create sub directories for each prompt and seed pair.
|
||||
|
||||
For example, if you provide the following prompts and seeds:
|
||||
|
||||
```
|
||||
prompts = ['a', 'b', 'c']
|
||||
prompts = ['a dog', 'a cat', 'a bird']
|
||||
seeds = [1, 2, 3]
|
||||
num_interpolation_steps = 5
|
||||
output_dir = 'output_dir'
|
||||
@ -722,7 +757,7 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
audio_offset = audio_start_sec + sum(num_interpolation_steps[:i]) / fps
|
||||
audio_duration = num_step / fps
|
||||
|
||||
self.generate_interpolation_clip(
|
||||
self.make_clip_frames(
|
||||
prompt_a,
|
||||
prompt_b,
|
||||
seed_a,
|
||||
@ -742,7 +777,8 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
offset=audio_offset,
|
||||
duration=audio_duration,
|
||||
fps=fps,
|
||||
margin=(1.0, 5.0),
|
||||
margin=margin,
|
||||
smooth=smooth,
|
||||
)
|
||||
if audio_filepath
|
||||
else None,
|
||||
@ -783,6 +819,23 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
embed = self.text_encoder(text_input.input_ids.to(self.device))[0]
|
||||
return embed
|
||||
|
||||
def init_noise(self, seed, noise_shape):
|
||||
"""Helper to initialize noise"""
|
||||
# randn does not exist on mps, so we create noise on CPU here and move it to the device after initialization
|
||||
if self.device.type == "mps":
|
||||
noise = torch.randn(
|
||||
noise_shape,
|
||||
device='cpu',
|
||||
generator=torch.Generator(device='cpu').manual_seed(seed),
|
||||
).to(self.device)
|
||||
else:
|
||||
noise = torch.randn(
|
||||
noise_shape,
|
||||
device=self.device,
|
||||
generator=torch.Generator(device=self.device).manual_seed(seed),
|
||||
)
|
||||
return noise
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, tiled=False, **kwargs):
|
||||
"""Same as diffusers `from_pretrained` but with tiled option, which makes images tilable"""
|
||||
@ -799,15 +852,6 @@ class StableDiffusionWalkPipeline(DiffusionPipeline):
|
||||
|
||||
patch_conv(padding_mode="circular")
|
||||
|
||||
return super().from_pretrained(*args, **kwargs)
|
||||
|
||||
|
||||
class NoCheck(ModelMixin):
|
||||
"""Can be used in place of safety checker. Use responsibly and at your own risk."""
|
||||
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.register_parameter(name="asdf", param=torch.nn.Parameter(torch.randn(3)))
|
||||
|
||||
def forward(self, images=None, **kwargs):
|
||||
return images, [False]
|
||||
pipeline = super().from_pretrained(*args, **kwargs)
|
||||
pipeline.tiled = tiled
|
||||
return pipeline
|
||||
@ -106,7 +106,7 @@ def stable_horde(outpath, prompt, seed, sampler_name, save_grid, batch_size,
|
||||
|
||||
log.append("Generating image with Stable Horde.")
|
||||
|
||||
st.session_state["progress_bar_text"].code('\n'.join(str(log)), language='')
|
||||
st.session_state["progress_bar_text"].code('\n'.join(log), language='')
|
||||
|
||||
# start time after garbage collection (or before?)
|
||||
start_time = time.time()
|
||||
@ -157,7 +157,7 @@ def stable_horde(outpath, prompt, seed, sampler_name, save_grid, batch_size,
|
||||
logger.debug(submit_results)
|
||||
|
||||
log.append(submit_results)
|
||||
st.session_state["progress_bar_text"].code('\n'.join(str(log)), language='')
|
||||
st.session_state["progress_bar_text"].code(''.join(str(log)), language='')
|
||||
|
||||
req_id = submit_results['id']
|
||||
is_done = False
|
||||
@ -282,7 +282,7 @@ def txt2img(prompt: str, ddim_steps: int, sampler_name: str, n_iter: int, batch_
|
||||
RealESRGAN_model: str = "RealESRGAN_x4plus_anime_6B", use_LDSR: bool = True, LDSR_model: str = "model",
|
||||
fp = None, variant_amount: float = 0.0,
|
||||
variant_seed: int = None, ddim_eta:float = 0.0, write_info_files:bool = True,
|
||||
use_stable_horde: bool = False, stable_horde_key:str = ''):
|
||||
use_stable_horde: bool = False, stable_horde_key:str = "0000000000"):
|
||||
|
||||
outpath = st.session_state['defaults'].general.outdir_txt2img
|
||||
|
||||
@ -410,7 +410,7 @@ def layout():
|
||||
sygil_suggestions.suggestion_area(placeholder)
|
||||
|
||||
# creating the page layout using columns
|
||||
col1, col2, col3 = st.columns([1,2,1], gap="large")
|
||||
col1, col2, col3 = st.columns([2,5,2], gap="large")
|
||||
|
||||
with col1:
|
||||
width = st.slider("Width:", min_value=st.session_state['defaults'].txt2img.width.min_value, max_value=st.session_state['defaults'].txt2img.width.max_value,
|
||||
@ -475,7 +475,7 @@ def layout():
|
||||
|
||||
with gallery_tab:
|
||||
st.session_state["gallery"] = st.empty()
|
||||
st.session_state["gallery"].info("Nothing but crickets here, try generating something first.")
|
||||
#st.session_state["gallery"].info("Nothing but crickets here, try generating something first.")
|
||||
|
||||
with col3:
|
||||
# If we have custom models available on the "models/custom"
|
||||
@ -502,7 +502,7 @@ def layout():
|
||||
with st.expander("Advanced"):
|
||||
with st.expander("Stable Horde"):
|
||||
use_stable_horde = st.checkbox("Use Stable Horde", value=False, help="Use the Stable Horde to generate images. More info can be found at https://stablehorde.net/")
|
||||
stable_horde_key = st.text_input("Stable Horde Api Key", value='', type="password",
|
||||
stable_horde_key = st.text_input("Stable Horde Api Key", value=st.session_state['defaults'].general.stable_horde_api, type="password",
|
||||
help="Optional Api Key used for the Stable Horde Bridge, if no api key is added the horde will be used anonymously.")
|
||||
|
||||
with st.expander("Output Settings"):
|
||||
@ -570,7 +570,9 @@ def layout():
|
||||
|
||||
#print (st.session_state["RealESRGAN_available"])
|
||||
st.session_state["upscaling_method"] = st.selectbox("Upscaling Method", upscaling_method_list,
|
||||
index=upscaling_method_list.index(str(st.session_state['defaults'].general.upscaling_method)))
|
||||
index=upscaling_method_list.index(st.session_state['defaults'].general.upscaling_method)
|
||||
if st.session_state['defaults'].general.upscaling_method in upscaling_method_list
|
||||
else 0)
|
||||
|
||||
if st.session_state["RealESRGAN_available"]:
|
||||
with st.expander("RealESRGAN"):
|
||||
@ -654,42 +656,9 @@ def layout():
|
||||
|
||||
message.success('Render Complete: ' + info + '; Stats: ' + stats, icon="✅")
|
||||
|
||||
#history_tab,col1,col2,col3,PlaceHolder,col1_cont,col2_cont,col3_cont = st.session_state['historyTab']
|
||||
|
||||
#if 'latestImages' in st.session_state:
|
||||
#for i in output_images:
|
||||
##push the new image to the list of latest images and remove the oldest one
|
||||
##remove the last index from the list\
|
||||
#st.session_state['latestImages'].pop()
|
||||
##add the new image to the start of the list
|
||||
#st.session_state['latestImages'].insert(0, i)
|
||||
#PlaceHolder.empty()
|
||||
#with PlaceHolder.container():
|
||||
#col1, col2, col3 = st.columns(3)
|
||||
#col1_cont = st.container()
|
||||
#col2_cont = st.container()
|
||||
#col3_cont = st.container()
|
||||
#images = st.session_state['latestImages']
|
||||
#with col1_cont:
|
||||
#with col1:
|
||||
#[st.image(images[index]) for index in [0, 3, 6] if index < len(images)]
|
||||
#with col2_cont:
|
||||
#with col2:
|
||||
#[st.image(images[index]) for index in [1, 4, 7] if index < len(images)]
|
||||
#with col3_cont:
|
||||
#with col3:
|
||||
#[st.image(images[index]) for index in [2, 5, 8] if index < len(images)]
|
||||
#historyGallery = st.empty()
|
||||
|
||||
## check if output_images length is the same as seeds length
|
||||
#with gallery_tab:
|
||||
#st.markdown(createHTMLGallery(output_images,seeds), unsafe_allow_html=True)
|
||||
|
||||
|
||||
#st.session_state['historyTab'] = [history_tab,col1,col2,col3,PlaceHolder,col1_cont,col2_cont,col3_cont]
|
||||
|
||||
with gallery_tab:
|
||||
logger.info(seeds)
|
||||
st.session_state["gallery"].text = ""
|
||||
sdGallery(output_images)
|
||||
|
||||
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@ -123,10 +123,16 @@ def layout():
|
||||
# specify the primary menu definition
|
||||
menu_data = [
|
||||
{'id': 'Stable Diffusion', 'label': 'Stable Diffusion', 'icon': 'bi bi-grid-1x2-fill'},
|
||||
{'id': 'Train','label':"Train", 'icon': "bi bi-lightbulb-fill", 'submenu':[
|
||||
{'id': 'Textual Inversion', 'label': 'Textual Inversion', 'icon': 'bi bi-lightbulb-fill'},
|
||||
{'id': 'Fine Tunning', 'label': 'Fine Tunning', 'icon': 'bi bi-lightbulb-fill'},
|
||||
]},
|
||||
{'id': 'Model Manager', 'label': 'Model Manager', 'icon': 'bi bi-cloud-arrow-down-fill'},
|
||||
#{'id': 'Tools','label':"Tools", 'icon': "bi bi-tools", 'submenu':[
|
||||
{'id': 'Tools','label':"Tools", 'icon': "bi bi-tools", 'submenu':[
|
||||
{'id': 'API Server', 'label': 'API Server', 'icon': 'bi bi-server'},
|
||||
{'id': 'Barfi/BaklavaJS', 'label': 'Barfi/BaklavaJS', 'icon': 'bi bi-diagram-3-fill'},
|
||||
#{'id': 'API Server', 'label': 'API Server', 'icon': 'bi bi-server'},
|
||||
]},
|
||||
{'id': 'Settings', 'label': 'Settings', 'icon': 'bi bi-gear-fill'},
|
||||
#{'icon': "fa-solid fa-radar",'label':"Dropdown1", 'submenu':[
|
||||
# {'id':' subid11','icon': "fa fa-paperclip", 'label':"Sub-item 1"},{'id':'subid12','icon': "💀", 'label':"Sub-item 2"},{'id':'subid13','icon': "fa fa-database", 'label':"Sub-item 3"}]},
|
||||
@ -172,6 +178,10 @@ def layout():
|
||||
#horizontal_orientation=False,
|
||||
#override_theme={'txc_inactive': 'white','menu_background':'#111', 'stVerticalBlock': '#111','txc_active':'yellow','option_active':'blue'})
|
||||
|
||||
#
|
||||
#if menu_id == "Home":
|
||||
#st.info("Under Construction. :construction_worker:")
|
||||
|
||||
if menu_id == "Stable Diffusion":
|
||||
# set the page url and title
|
||||
#st.experimental_set_query_params(page='stable-diffusion')
|
||||
@ -180,9 +190,10 @@ def layout():
|
||||
except NameError:
|
||||
st.experimental_rerun()
|
||||
|
||||
txt2img_tab, img2img_tab, txt2vid_tab, img2txt_tab, concept_library_tab = st.tabs(["Text-to-Image", "Image-to-Image",
|
||||
txt2img_tab, img2img_tab, txt2vid_tab, img2txt_tab, post_processing_tab, concept_library_tab = st.tabs(["Text-to-Image", "Image-to-Image",
|
||||
#"Inpainting",
|
||||
"Text-to-Video", "Image-To-Text",
|
||||
"Concept Library"])
|
||||
"Post-Processing","Concept Library"])
|
||||
#with home_tab:
|
||||
#from home import layout
|
||||
#layout()
|
||||
@ -207,6 +218,10 @@ def layout():
|
||||
from img2txt import layout
|
||||
layout()
|
||||
|
||||
with post_processing_tab:
|
||||
from post_processing import layout
|
||||
layout()
|
||||
|
||||
with concept_library_tab:
|
||||
from sd_concept_library import layout
|
||||
layout()
|
||||
@ -222,11 +237,21 @@ def layout():
|
||||
from textual_inversion import layout
|
||||
layout()
|
||||
|
||||
elif menu_id == 'Fine Tunning':
|
||||
#from textual_inversion import layout
|
||||
#layout()
|
||||
st.info("Under Construction. :construction_worker:")
|
||||
|
||||
elif menu_id == 'API Server':
|
||||
set_page_title("API Server - Stable Diffusion Playground")
|
||||
from APIServer import layout
|
||||
layout()
|
||||
|
||||
elif menu_id == 'Barfi/BaklavaJS':
|
||||
set_page_title("Barfi/BaklavaJS - Stable Diffusion Playground")
|
||||
from barfi_baklavajs import layout
|
||||
layout()
|
||||
|
||||
elif menu_id == 'Settings':
|
||||
set_page_title("Settings - Stable Diffusion Playground")
|
||||
|
||||
|
||||
@ -103,10 +103,10 @@ call "%v_conda_path%\Scripts\activate.bat" "%v_conda_env_name%"
|
||||
:PROMPT
|
||||
set SETUPTOOLS_USE_DISTUTILS=stdlib
|
||||
IF EXIST "models\ldm\stable-diffusion-v1\Stable Diffusion v1.5.ckpt" (
|
||||
python -m streamlit run scripts\webui_streamlit.py --theme.base dark --server.address localhost
|
||||
python -m streamlit run scripts\webui_streamlit.py --theme.base dark
|
||||
) ELSE (
|
||||
echo Your model file does not exist! Once the WebUI launches please visit the Model Manager page and download the models by using the Download button for each model.
|
||||
python -m streamlit run scripts\webui_streamlit.py --theme.base dark --server.address localhost
|
||||
python -m streamlit run scripts\webui_streamlit.py --theme.base dark
|
||||
)
|
||||
|
||||
::cmd /k
|
||||
|
||||
4
webui.sh
4
webui.sh
@ -156,12 +156,12 @@ launch_webui () {
|
||||
done
|
||||
printf "\n\n########## LAUNCH USING STREAMLIT OR GRADIO? ##########\n\n"
|
||||
printf "Do you wish to run the WebUI using the Gradio or StreamLit Interface?\n\n"
|
||||
printf "Streamlit: \nHas A More Modern UI \nMore Features Planned \nWill Be The Main UI Going Forward \nCurrently In Active Development \nMissing Some Gradio Features\n\n"
|
||||
printf "Streamlit: \nHas A More Modern UI \nMore Features Planned \nWill Be The Main UI Going Forward \nCurrently In Active Development \n\n"
|
||||
printf "Gradio: \nCurrently Feature Complete \nUses An Older Interface Style \nWill Not Receive Major Updates\n\n"
|
||||
printf "Which Version of the WebUI Interface do you wish to use?\n"
|
||||
select yn in "Streamlit" "Gradio"; do
|
||||
case $yn in
|
||||
Streamlit ) printf "\nStarting Stable Diffusion WebUI: Streamlit Interface. Please Wait...\n"; python -m streamlit run scripts/webui_streamlit.py --theme.base dark --server.address localhost; break;;
|
||||
Streamlit ) printf "\nStarting Stable Diffusion WebUI: Streamlit Interface. Please Wait...\n"; python -m streamlit run scripts/webui_streamlit.py; break;;
|
||||
Gradio ) printf "\nStarting Stable Diffusion WebUI: Gradio Interface. Please Wait...\n"; python scripts/relauncher.py "$@"; break;;
|
||||
esac
|
||||
done
|
||||
|
||||
Loading…
Reference in New Issue
Block a user