mirror of
https://github.com/Sygil-Dev/sygil-webui.git
synced 2024-12-14 22:13:41 +03:00
The webui_streamlit.py file has been split into multiple modules containing their own code making it easier to work with than a single big file.
The list of modules is as follow: - webuit_streamlit.py: contains the main layout as well as the functions that load the css which is needed by the layout. - webui_streamlit_old.py: contains the code for the previous version of the WebUI. Will be removed once the new UI code starts to get used and if everything works as it should. - txt2img.py: contains the code for the txt2img tab. - img2img.py: contains the code for the img2img tab. - txt2vid.py: contains the code for the txt2vid tab. - sd_utils.py: contains utility functions used by more than one module, any function that meets such condition should be placed here. - ModelManager.py: contains the code for the Model Manager page on the sidebar menu. - Settings.py: contains the code for the Settings page on the sidebar menu. - home.py: contains the code for the Home tab, history and gallery implemented by @devilismyfriend. - imglab.py: contains the code for the Image Lab tab implemented by @devilismyfriend
This commit is contained in:
ZeroCool940711
parent
76de9f9b65
commit
ede343a269
@ -58,8 +58,8 @@ button[kind="header"] {
|
||||
|
||||
/* added to avoid main sectors (all element to the right of sidebar from) moving */
|
||||
section[data-testid="stSidebar"] {
|
||||
width: 3% !important;
|
||||
min-width: 3% !important;
|
||||
width: 3.5% !important;
|
||||
min-width: 3.5% !important;
|
||||
}
|
||||
|
||||
/* The navigation menu specs and size */
|
||||
|
||||
32
scripts/ModelManager.py
Normal file
32
scripts/ModelManager.py
Normal file
@ -0,0 +1,32 @@
|
||||
from webui_streamlit import st
|
||||
|
||||
from sd_utils import *
|
||||
|
||||
def layout():
|
||||
#search = st.text_input(label="Search", placeholder="Type the name of the model you want to search for.", help="")
|
||||
|
||||
csvString = f"""
|
||||
,Stable Diffusion v1.4 , ./models/ldm/stable-diffusion-v1 , https://www.googleapis.com/storage/v1/b/aai-blog-files/o/sd-v1-4.ckpt?alt=media
|
||||
,GFPGAN v1.3 , ./src/gfpgan/experiments/pretrained_models , https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth
|
||||
,RealESRGAN_x4plus , ./src/realesrgan/experiments/pretrained_models , https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth
|
||||
,RealESRGAN_x4plus_anime_6B , ./src/realesrgan/experiments/pretrained_models , https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth
|
||||
,Waifu Diffusion v1.2 , ./models/custom , http://wd.links.sd:8880/wd-v1-2-full-ema.ckpt
|
||||
,TrinArt Stable Diffusion v2 , ./models/custom , https://huggingface.co/naclbit/trinart_stable_diffusion_v2/resolve/main/trinart2_step115000.ckpt
|
||||
"""
|
||||
colms = st.columns((1, 3, 5, 5))
|
||||
columns = ["№",'Model Name','Save Location','Download Link']
|
||||
|
||||
# Convert String into StringIO
|
||||
csvStringIO = StringIO(csvString)
|
||||
df = pd.read_csv(csvStringIO, sep=",", header=None, names=columns)
|
||||
|
||||
for col, field_name in zip(colms, columns):
|
||||
# table header
|
||||
col.write(field_name)
|
||||
|
||||
for x, model_name in enumerate(df["Model Name"]):
|
||||
col1, col2, col3, col4 = st.columns((1, 3, 4, 6))
|
||||
col1.write(x) # index
|
||||
col2.write(df['Model Name'][x])
|
||||
col3.write(df['Save Location'][x])
|
||||
col4.write(df['Download Link'][x])
|
||||
142
scripts/home.py
Normal file
142
scripts/home.py
Normal file
@ -0,0 +1,142 @@
|
||||
from webui_streamlit import st
|
||||
from sd_utils import *
|
||||
import os
|
||||
from PIL import Image
|
||||
|
||||
try:
|
||||
# this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start.
|
||||
from transformers import logging
|
||||
|
||||
logging.set_verbosity_error()
|
||||
except:
|
||||
pass
|
||||
|
||||
class plugin_info():
|
||||
plugname = "home"
|
||||
description = "Home"
|
||||
isTab = True
|
||||
displayPriority = 0
|
||||
|
||||
def getLatestGeneratedImagesFromPath():
|
||||
#get the latest images from the generated images folder
|
||||
#get the path to the generated images folder
|
||||
generatedImagesPath = os.path.join(os.getcwd(),'outputs')
|
||||
#get all the files from the folders and subfolders
|
||||
files = []
|
||||
#get the laest 10 images from the output folder without walking the subfolders
|
||||
for r, d, f in os.walk(generatedImagesPath):
|
||||
for file in f:
|
||||
if '.png' in file:
|
||||
files.append(os.path.join(r, file))
|
||||
#sort the files by date
|
||||
files.sort(key=os.path.getmtime)
|
||||
#reverse the list so the latest images are first
|
||||
for f in files:
|
||||
img = Image.open(f)
|
||||
files[files.index(f)] = img
|
||||
#get the latest 10 files
|
||||
#get all the files with the .png or .jpg extension
|
||||
#sort files by date
|
||||
#get the latest 10 files
|
||||
latestFiles = files[-10:]
|
||||
#reverse the list
|
||||
latestFiles.reverse()
|
||||
return latestFiles
|
||||
|
||||
def get_images_from_lexica():
|
||||
#scrape images from lexica.art
|
||||
#get the html from the page
|
||||
#get the html with cookies and javascript
|
||||
apiEndpoint = r'https://lexica.art/api/trpc/prompts.infinitePrompts?batch=1&input=%7B%220%22%3A%7B%22json%22%3A%7B%22limit%22%3A10%2C%22text%22%3A%22%22%2C%22cursor%22%3A10%7D%7D%7D'
|
||||
#REST API call
|
||||
#
|
||||
from requests_html import HTMLSession
|
||||
session = HTMLSession()
|
||||
|
||||
response = session.get(apiEndpoint)
|
||||
#req = requests.Session()
|
||||
#req.headers['user-agent'] = 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/96.0.4664.45 Safari/537.36'
|
||||
#response = req.get(apiEndpoint)
|
||||
print(response.status_code)
|
||||
print(response.text)
|
||||
#get the json from the response
|
||||
#json = response.json()
|
||||
#get the prompts from the json
|
||||
print(response)
|
||||
#session = requests.Session()
|
||||
#parseEndpointJson = session.get(apiEndpoint,headers=headers,verify=False)
|
||||
#print(parseEndpointJson)
|
||||
#print('test2')
|
||||
#page = requests.get("https://lexica.art/", headers={'User-Agent': 'Mozilla/5.0'})
|
||||
#parse the html
|
||||
#soup = BeautifulSoup(page.content, 'html.parser')
|
||||
#find all the images
|
||||
#print(soup)
|
||||
#images = soup.find_all('alt-image')
|
||||
#create a list to store the image urls
|
||||
image_urls = []
|
||||
#loop through the images
|
||||
for image in images:
|
||||
#get the url
|
||||
image_url = image['src']
|
||||
#add it to the list
|
||||
image_urls.append('http://www.lexica.art/'+image_url)
|
||||
#return the list
|
||||
print(image_urls)
|
||||
return image_urls
|
||||
|
||||
def layout():
|
||||
#streamlit home page layout
|
||||
#center the title
|
||||
st.markdown("<h1 style='text-align: center; color: white;'>Welcome, let's make some 🎨</h1>", unsafe_allow_html=True)
|
||||
#make a gallery of images
|
||||
#st.markdown("<h2 style='text-align: center; color: white;'>Gallery</h2>", unsafe_allow_html=True)
|
||||
#create a gallery of images using columns
|
||||
#col1, col2, col3 = st.columns(3)
|
||||
#load the images
|
||||
#create 3 columns
|
||||
# create a tab for the gallery
|
||||
#st.markdown("<h2 style='text-align: center; color: white;'>Gallery</h2>", unsafe_allow_html=True)
|
||||
#st.markdown("<h2 style='text-align: center; color: white;'>Gallery</h2>", unsafe_allow_html=True)
|
||||
history_tab, discover_tabs, settings_tab = st.tabs(["History","Discover","Settings"])
|
||||
with discover_tabs:
|
||||
st.markdown("<h1 style='text-align: center; color: white;'>Soon :)</h1>", unsafe_allow_html=True)
|
||||
with settings_tab:
|
||||
st.markdown("<h1 style='text-align: center; color: white;'>Soon :)</h1>", unsafe_allow_html=True)
|
||||
with history_tab:
|
||||
placeholder = st.empty()
|
||||
|
||||
|
||||
|
||||
latestImages = getLatestGeneratedImagesFromPath()
|
||||
st.session_state['latestImages'] = latestImages
|
||||
|
||||
#populate the 3 images per column
|
||||
with placeholder.container():
|
||||
col1, col2, col3 = st.columns(3)
|
||||
col1_cont = st.container()
|
||||
col2_cont = st.container()
|
||||
col3_cont = st.container()
|
||||
with col1_cont:
|
||||
with col1:
|
||||
st.image(st.session_state['latestImages'][0])
|
||||
st.image(st.session_state['latestImages'][3])
|
||||
st.image(st.session_state['latestImages'][6])
|
||||
with col2_cont:
|
||||
with col2:
|
||||
st.image(st.session_state['latestImages'][1])
|
||||
st.image(st.session_state['latestImages'][4])
|
||||
st.image(st.session_state['latestImages'][7])
|
||||
with col3_cont:
|
||||
with col3:
|
||||
st.image(st.session_state['latestImages'][2])
|
||||
st.image(st.session_state['latestImages'][5])
|
||||
st.image(st.session_state['latestImages'][8])
|
||||
st.session_state['historyTab'] = [history_tab,col1,col2,col3,placeholder,col1_cont,col2_cont,col3_cont]
|
||||
#display the images
|
||||
#add a button to the gallery
|
||||
#st.markdown("<h2 style='text-align: center; color: white;'>Try it out</h2>", unsafe_allow_html=True)
|
||||
#create a button to the gallery
|
||||
#if st.button("Try it out"):
|
||||
#if the button is clicked, go to the gallery
|
||||
#st.experimental_rerun()
|
||||
569
scripts/img2img.py
Normal file
569
scripts/img2img.py
Normal file
@ -0,0 +1,569 @@
|
||||
from webui_streamlit import st
|
||||
from sd_utils import *
|
||||
|
||||
from streamlit import StopException
|
||||
|
||||
from PIL import Image, ImageOps
|
||||
import torch
|
||||
import k_diffusion as K
|
||||
import numpy as np
|
||||
import time
|
||||
import torch
|
||||
import skimage
|
||||
from ldm.models.diffusion.ddim import DDIMSampler
|
||||
from ldm.models.diffusion.plms import PLMSSampler
|
||||
|
||||
try:
|
||||
# this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start.
|
||||
from transformers import logging
|
||||
|
||||
logging.set_verbosity_error()
|
||||
except:
|
||||
pass
|
||||
|
||||
def img2img(prompt: str = '', init_info: any = None, init_info_mask: any = None, mask_mode: int = 0, mask_blur_strength: int = 3,
|
||||
mask_restore: bool = False, ddim_steps: int = 50, sampler_name: str = 'DDIM',
|
||||
n_iter: int = 1, cfg_scale: float = 7.5, denoising_strength: float = 0.8,
|
||||
seed: int = -1, noise_mode: int = 0, find_noise_steps: str = "", height: int = 512, width: int = 512, resize_mode: int = 0, fp = None,
|
||||
variant_amount: float = None, variant_seed: int = None, ddim_eta:float = 0.0,
|
||||
write_info_files:bool = True, RealESRGAN_model: str = "RealESRGAN_x4plus_anime_6B",
|
||||
separate_prompts:bool = False, normalize_prompt_weights:bool = True,
|
||||
save_individual_images: bool = True, save_grid: bool = True, group_by_prompt: bool = True,
|
||||
save_as_jpg: bool = True, use_GFPGAN: bool = True, use_RealESRGAN: bool = True, loopback: bool = False,
|
||||
random_seed_loopback: bool = False
|
||||
):
|
||||
|
||||
outpath = defaults.general.outdir_img2img or defaults.general.outdir or "outputs/img2img-samples"
|
||||
#err = False
|
||||
#loopback = False
|
||||
#skip_save = False
|
||||
seed = seed_to_int(seed)
|
||||
|
||||
batch_size = 1
|
||||
|
||||
#prompt_matrix = 0
|
||||
#normalize_prompt_weights = 1 in toggles
|
||||
#loopback = 2 in toggles
|
||||
#random_seed_loopback = 3 in toggles
|
||||
#skip_save = 4 not in toggles
|
||||
#save_grid = 5 in toggles
|
||||
#sort_samples = 6 in toggles
|
||||
#write_info_files = 7 in toggles
|
||||
#write_sample_info_to_log_file = 8 in toggles
|
||||
#jpg_sample = 9 in toggles
|
||||
#use_GFPGAN = 10 in toggles
|
||||
#use_RealESRGAN = 11 in toggles
|
||||
|
||||
if sampler_name == 'PLMS':
|
||||
sampler = PLMSSampler(st.session_state["model"])
|
||||
elif sampler_name == 'DDIM':
|
||||
sampler = DDIMSampler(st.session_state["model"])
|
||||
elif sampler_name == 'k_dpm_2_a':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'dpm_2_ancestral')
|
||||
elif sampler_name == 'k_dpm_2':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'dpm_2')
|
||||
elif sampler_name == 'k_euler_a':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'euler_ancestral')
|
||||
elif sampler_name == 'k_euler':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'euler')
|
||||
elif sampler_name == 'k_heun':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'heun')
|
||||
elif sampler_name == 'k_lms':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'lms')
|
||||
else:
|
||||
raise Exception("Unknown sampler: " + sampler_name)
|
||||
|
||||
def process_init_mask(init_mask: Image):
|
||||
if init_mask.mode == "RGBA":
|
||||
init_mask = init_mask.convert('RGBA')
|
||||
background = Image.new('RGBA', init_mask.size, (0, 0, 0))
|
||||
init_mask = Image.alpha_composite(background, init_mask)
|
||||
init_mask = init_mask.convert('RGB')
|
||||
return init_mask
|
||||
|
||||
init_img = init_info
|
||||
init_mask = None
|
||||
if mask_mode == 0:
|
||||
if init_info_mask:
|
||||
init_mask = process_init_mask(init_info_mask)
|
||||
elif mask_mode == 1:
|
||||
if init_info_mask:
|
||||
init_mask = process_init_mask(init_info_mask)
|
||||
init_mask = ImageOps.invert(init_mask)
|
||||
elif mask_mode == 2:
|
||||
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 = 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 == 2 or noise_mode == 3) 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(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 = matched_noise.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 = Image.fromarray(np.clip(noised * 255., 0., 255.).astype(np.uint8), mode="RGB")
|
||||
st.session_state["editor_image"].image(init_img) # debug
|
||||
|
||||
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 = 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(st.session_state["device"])
|
||||
|
||||
if defaults.general.optimized:
|
||||
modelFS.to(st.session_state["device"] )
|
||||
|
||||
init_image = 2. * image - 1.
|
||||
init_image = init_image.to(st.session_state["device"])
|
||||
init_latent = (st.session_state["model"] if not defaults.general.optimized else modelFS).get_first_stage_encoding((st.session_state["model"] if not defaults.general.optimized else modelFS).encode_first_stage(init_image)) # move to latent space
|
||||
|
||||
if defaults.general.optimized:
|
||||
mem = torch.cuda.memory_allocated()/1e6
|
||||
modelFS.to("cpu")
|
||||
while(torch.cuda.memory_allocated()/1e6 >= mem):
|
||||
time.sleep(1)
|
||||
|
||||
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,
|
||||
callback=generation_callback)
|
||||
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(st.session_state["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
|
||||
|
||||
|
||||
|
||||
if loopback:
|
||||
output_images, info = None, None
|
||||
history = []
|
||||
initial_seed = None
|
||||
|
||||
do_color_correction = False
|
||||
try:
|
||||
from skimage import exposure
|
||||
do_color_correction = True
|
||||
except:
|
||||
print("Install scikit-image to perform color correction on loopback")
|
||||
|
||||
for i in range(n_iter):
|
||||
if do_color_correction and i == 0:
|
||||
correction_target = cv2.cvtColor(np.asarray(init_img.copy()), cv2.COLOR_RGB2LAB)
|
||||
|
||||
output_images, seed, info, stats = process_images(
|
||||
outpath=outpath,
|
||||
func_init=init,
|
||||
func_sample=sample,
|
||||
prompt=prompt,
|
||||
seed=seed,
|
||||
sampler_name=sampler_name,
|
||||
save_grid=save_grid,
|
||||
batch_size=1,
|
||||
n_iter=1,
|
||||
steps=ddim_steps,
|
||||
cfg_scale=cfg_scale,
|
||||
width=width,
|
||||
height=height,
|
||||
prompt_matrix=separate_prompts,
|
||||
use_GFPGAN=use_GFPGAN,
|
||||
use_RealESRGAN=use_RealESRGAN, # Forcefully disable upscaling when using loopback
|
||||
realesrgan_model_name=RealESRGAN_model,
|
||||
fp=fp,
|
||||
normalize_prompt_weights=normalize_prompt_weights,
|
||||
save_individual_images=save_individual_images,
|
||||
init_img=init_img,
|
||||
init_mask=init_mask,
|
||||
mask_blur_strength=mask_blur_strength,
|
||||
mask_restore=mask_restore,
|
||||
denoising_strength=denoising_strength,
|
||||
noise_mode=noise_mode,
|
||||
find_noise_steps=find_noise_steps,
|
||||
resize_mode=resize_mode,
|
||||
uses_loopback=loopback,
|
||||
uses_random_seed_loopback=random_seed_loopback,
|
||||
sort_samples=group_by_prompt,
|
||||
write_info_files=write_info_files,
|
||||
jpg_sample=save_as_jpg
|
||||
)
|
||||
|
||||
if initial_seed is None:
|
||||
initial_seed = seed
|
||||
|
||||
init_img = output_images[0]
|
||||
|
||||
if do_color_correction and correction_target is not None:
|
||||
init_img = Image.fromarray(cv2.cvtColor(exposure.match_histograms(
|
||||
cv2.cvtColor(
|
||||
np.asarray(init_img),
|
||||
cv2.COLOR_RGB2LAB
|
||||
),
|
||||
correction_target,
|
||||
channel_axis=2
|
||||
), cv2.COLOR_LAB2RGB).astype("uint8"))
|
||||
|
||||
if not random_seed_loopback:
|
||||
seed = seed + 1
|
||||
else:
|
||||
seed = seed_to_int(None)
|
||||
|
||||
denoising_strength = max(denoising_strength * 0.95, 0.1)
|
||||
history.append(init_img)
|
||||
|
||||
output_images = history
|
||||
seed = initial_seed
|
||||
|
||||
else:
|
||||
output_images, seed, info, stats = process_images(
|
||||
outpath=outpath,
|
||||
func_init=init,
|
||||
func_sample=sample,
|
||||
prompt=prompt,
|
||||
seed=seed,
|
||||
sampler_name=sampler_name,
|
||||
save_grid=save_grid,
|
||||
batch_size=batch_size,
|
||||
n_iter=n_iter,
|
||||
steps=ddim_steps,
|
||||
cfg_scale=cfg_scale,
|
||||
width=width,
|
||||
height=height,
|
||||
prompt_matrix=separate_prompts,
|
||||
use_GFPGAN=use_GFPGAN,
|
||||
use_RealESRGAN=use_RealESRGAN,
|
||||
realesrgan_model_name=RealESRGAN_model,
|
||||
fp=fp,
|
||||
normalize_prompt_weights=normalize_prompt_weights,
|
||||
save_individual_images=save_individual_images,
|
||||
init_img=init_img,
|
||||
init_mask=init_mask,
|
||||
mask_blur_strength=mask_blur_strength,
|
||||
denoising_strength=denoising_strength,
|
||||
noise_mode=noise_mode,
|
||||
find_noise_steps=find_noise_steps,
|
||||
mask_restore=mask_restore,
|
||||
resize_mode=resize_mode,
|
||||
uses_loopback=loopback,
|
||||
sort_samples=group_by_prompt,
|
||||
write_info_files=write_info_files,
|
||||
jpg_sample=save_as_jpg
|
||||
)
|
||||
|
||||
del sampler
|
||||
|
||||
return output_images, seed, info, stats
|
||||
|
||||
#
|
||||
|
||||
|
||||
def layout():
|
||||
with st.form("img2img-inputs"):
|
||||
st.session_state["generation_mode"] = "img2img"
|
||||
|
||||
img2img_input_col, img2img_generate_col = st.columns([10,1])
|
||||
with img2img_input_col:
|
||||
#prompt = st.text_area("Input Text","")
|
||||
prompt = st.text_input("Input Text","", placeholder="A corgi wearing a top hat as an oil painting.")
|
||||
|
||||
# Every form must have a submit button, the extra blank spaces is a temp way to align it with the input field. Needs to be done in CSS or some other way.
|
||||
img2img_generate_col.write("")
|
||||
img2img_generate_col.write("")
|
||||
generate_button = img2img_generate_col.form_submit_button("Generate")
|
||||
|
||||
|
||||
# creating the page layout using columns
|
||||
col1_img2img_layout, col2_img2img_layout, col3_img2img_layout = st.columns([1,2,2], gap="small")
|
||||
|
||||
with col1_img2img_layout:
|
||||
# If we have custom models available on the "models/custom"
|
||||
#folder then we show a menu to select which model we want to use, otherwise we use the main model for SD
|
||||
if st.session_state["CustomModel_available"]:
|
||||
st.session_state["custom_model"] = st.selectbox("Custom Model:", st.session_state["custom_models"],
|
||||
index=st.session_state["custom_models"].index(defaults.general.default_model),
|
||||
help="Select the model you want to use. This option is only available if you have custom models \
|
||||
on your 'models/custom' folder. The model name that will be shown here is the same as the name\
|
||||
the file for the model has on said folder, it is recommended to give the .ckpt file a name that \
|
||||
will make it easier for you to distinguish it from other models. Default: Stable Diffusion v1.4")
|
||||
else:
|
||||
st.session_state["custom_model"] = "Stable Diffusion v1.4"
|
||||
|
||||
|
||||
st.session_state["sampling_steps"] = st.slider("Sampling Steps", value=defaults.img2img.sampling_steps, min_value=1, max_value=500)
|
||||
|
||||
sampler_name_list = ["k_lms", "k_euler", "k_euler_a", "k_dpm_2", "k_dpm_2_a", "k_heun", "PLMS", "DDIM"]
|
||||
st.session_state["sampler_name"] = st.selectbox("Sampling method",sampler_name_list,
|
||||
index=sampler_name_list.index(defaults.img2img.sampler_name), help="Sampling method to use.")
|
||||
|
||||
mask_mode_list = ["Mask", "Inverted mask", "Image alpha"]
|
||||
mask_mode = st.selectbox("Mask Mode", mask_mode_list,
|
||||
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."
|
||||
)
|
||||
mask_mode = mask_mode_list.index(mask_mode)
|
||||
|
||||
width = st.slider("Width:", min_value=64, max_value=1024, value=defaults.img2img.width, step=64)
|
||||
height = st.slider("Height:", min_value=64, max_value=1024, value=defaults.img2img.height, step=64)
|
||||
seed = st.text_input("Seed:", value=defaults.img2img.seed, help=" The seed to use, if left blank a random seed will be generated.")
|
||||
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)
|
||||
find_noise_steps = st.slider("Find Noise Steps", value=100, min_value=1, max_value=500)
|
||||
batch_count = st.slider("Batch count.", min_value=1, max_value=100, value=defaults.img2img.batch_count, step=1,
|
||||
help="How many iterations or batches of images to generate in total.")
|
||||
|
||||
#
|
||||
with st.expander("Advanced"):
|
||||
separate_prompts = st.checkbox("Create Prompt Matrix.", value=defaults.img2img.separate_prompts,
|
||||
help="Separate multiple prompts using the `|` character, and get all combinations of them.")
|
||||
normalize_prompt_weights = st.checkbox("Normalize Prompt Weights.", value=defaults.img2img.normalize_prompt_weights,
|
||||
help="Ensure the sum of all weights add up to 1.0")
|
||||
loopback = st.checkbox("Loopback.", value=defaults.img2img.loopback, help="Use images from previous batch when creating next batch.")
|
||||
random_seed_loopback = st.checkbox("Random loopback seed.", value=defaults.img2img.random_seed_loopback, help="Random loopback seed")
|
||||
save_individual_images = st.checkbox("Save individual images.", value=defaults.img2img.save_individual_images,
|
||||
help="Save each image generated before any filter or enhancement is applied.")
|
||||
save_grid = st.checkbox("Save grid",value=defaults.img2img.save_grid, help="Save a grid with all the images generated into a single image.")
|
||||
group_by_prompt = st.checkbox("Group results by prompt", value=defaults.img2img.group_by_prompt,
|
||||
help="Saves all the images with the same prompt into the same folder. \
|
||||
When using a prompt matrix each prompt combination will have its own folder.")
|
||||
write_info_files = st.checkbox("Write Info file", value=defaults.img2img.write_info_files,
|
||||
help="Save a file next to the image with informartion about the generation.")
|
||||
save_as_jpg = st.checkbox("Save samples as jpg", value=defaults.img2img.save_as_jpg, help="Saves the images as jpg instead of png.")
|
||||
|
||||
if st.session_state["GFPGAN_available"]:
|
||||
use_GFPGAN = st.checkbox("Use GFPGAN", value=defaults.img2img.use_GFPGAN, help="Uses the GFPGAN model to improve faces after the generation.\
|
||||
This greatly improve the quality and consistency of faces but uses extra VRAM. Disable if you need the extra VRAM.")
|
||||
else:
|
||||
use_GFPGAN = False
|
||||
|
||||
if st.session_state["RealESRGAN_available"]:
|
||||
st.session_state["use_RealESRGAN"] = st.checkbox("Use RealESRGAN", value=defaults.img2img.use_RealESRGAN,
|
||||
help="Uses the RealESRGAN model to upscale the images after the generation.\
|
||||
This greatly improve the quality and lets you have high resolution images but uses extra VRAM. Disable if you need the extra VRAM.")
|
||||
st.session_state["RealESRGAN_model"] = st.selectbox("RealESRGAN model", ["RealESRGAN_x4plus", "RealESRGAN_x4plus_anime_6B"], index=0)
|
||||
else:
|
||||
st.session_state["use_RealESRGAN"] = False
|
||||
st.session_state["RealESRGAN_model"] = "RealESRGAN_x4plus"
|
||||
|
||||
variant_amount = st.slider("Variant Amount:", value=defaults.img2img.variant_amount, min_value=0.0, max_value=1.0, step=0.01)
|
||||
variant_seed = st.text_input("Variant Seed:", value=defaults.img2img.variant_seed,
|
||||
help="The seed to use when generating a variant, if left blank a random seed will be generated.")
|
||||
cfg_scale = st.slider("CFG (Classifier Free Guidance Scale):", min_value=1.0, max_value=30.0, value=defaults.img2img.cfg_scale, step=0.5,
|
||||
help="How strongly the image should follow the prompt.")
|
||||
batch_size = st.slider("Batch size", min_value=1, max_value=100, value=defaults.img2img.batch_size, step=1,
|
||||
help="How many images are at once in a batch.\
|
||||
It increases the VRAM usage a lot but if you have enough VRAM it can reduce the time it takes to finish \
|
||||
generation as more images are generated at once.\
|
||||
Default: 1")
|
||||
|
||||
st.session_state["denoising_strength"] = st.slider("Denoising Strength:", value=defaults.img2img.denoising_strength,
|
||||
min_value=0.01, max_value=1.0, step=0.01)
|
||||
|
||||
with st.expander("Preview Settings"):
|
||||
st.session_state["update_preview"] = st.checkbox("Update Image Preview", value=defaults.img2img.update_preview,
|
||||
help="If enabled the image preview will be updated during the generation instead of at the end. \
|
||||
You can use the Update Preview \Frequency option bellow to customize how frequent it's updated. \
|
||||
By default this is enabled and the frequency is set to 1 step.")
|
||||
|
||||
st.session_state["update_preview_frequency"] = st.text_input("Update Image Preview Frequency", value=defaults.img2img.update_preview_frequency,
|
||||
help="Frequency in steps at which the the preview image is updated. By default the frequency \
|
||||
is set to 1 step.")
|
||||
|
||||
with col2_img2img_layout:
|
||||
editor_tab = st.tabs(["Editor"])
|
||||
|
||||
editor_image = st.empty()
|
||||
st.session_state["editor_image"] = editor_image
|
||||
|
||||
st.form_submit_button("Refresh")
|
||||
|
||||
masked_image_holder = st.empty()
|
||||
image_holder = st.empty()
|
||||
|
||||
uploaded_images = st.file_uploader(
|
||||
"Upload Image", accept_multiple_files=False, type=["png", "jpg", "jpeg"],
|
||||
help="Upload an image which will be used for the image to image generation.",
|
||||
)
|
||||
if uploaded_images:
|
||||
image = Image.open(uploaded_images).convert('RGBA')
|
||||
new_img = image.resize((width, height))
|
||||
image_holder.image(new_img)
|
||||
|
||||
mask_holder = st.empty()
|
||||
|
||||
uploaded_masks = st.file_uploader(
|
||||
"Upload Mask", accept_multiple_files=False, type=["png", "jpg", "jpeg"],
|
||||
help="Upload an mask image which will be used for masking the image to image generation.",
|
||||
)
|
||||
if uploaded_masks:
|
||||
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)
|
||||
|
||||
final_img.putalpha(alpha_layer)
|
||||
|
||||
with masked_image_holder.container():
|
||||
st.text("Masked Image Preview")
|
||||
st.image(final_img)
|
||||
|
||||
|
||||
with col3_img2img_layout:
|
||||
result_tab = st.tabs(["Result"])
|
||||
|
||||
# create an empty container for the image, progress bar, etc so we can update it later and use session_state to hold them globally.
|
||||
preview_image = st.empty()
|
||||
st.session_state["preview_image"] = preview_image
|
||||
|
||||
#st.session_state["loading"] = st.empty()
|
||||
|
||||
st.session_state["progress_bar_text"] = st.empty()
|
||||
st.session_state["progress_bar"] = st.empty()
|
||||
|
||||
|
||||
message = st.empty()
|
||||
|
||||
#if uploaded_images:
|
||||
#image = Image.open(uploaded_images).convert('RGB')
|
||||
##img_array = np.array(image) # if you want to pass it to OpenCV
|
||||
#new_img = image.resize((width, height))
|
||||
#st.image(new_img, use_column_width=True)
|
||||
|
||||
|
||||
if generate_button:
|
||||
#print("Loading models")
|
||||
# load the models when we hit the generate button for the first time, it wont be loaded after that so dont worry.
|
||||
load_models(False, use_GFPGAN, st.session_state["use_RealESRGAN"], st.session_state["RealESRGAN_model"], st.session_state["CustomModel_available"],
|
||||
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
|
||||
new_mask = None
|
||||
if uploaded_masks:
|
||||
mask = Image.open(uploaded_masks).convert('RGBA')
|
||||
new_mask = mask.resize((width, height))
|
||||
|
||||
try:
|
||||
output_images, seed, info, stats = img2img(prompt=prompt, init_info=new_img, init_info_mask=new_mask, mask_mode=mask_mode,
|
||||
ddim_steps=st.session_state["sampling_steps"],
|
||||
sampler_name=st.session_state["sampler_name"], n_iter=batch_count,
|
||||
cfg_scale=cfg_scale, denoising_strength=st.session_state["denoising_strength"], variant_seed=variant_seed,
|
||||
seed=seed, noise_mode=noise_mode, find_noise_steps=find_noise_steps, width=width,
|
||||
height=height, fp=defaults.general.fp, variant_amount=variant_amount,
|
||||
ddim_eta=0.0, write_info_files=write_info_files, RealESRGAN_model=st.session_state["RealESRGAN_model"],
|
||||
separate_prompts=separate_prompts, normalize_prompt_weights=normalize_prompt_weights,
|
||||
save_individual_images=save_individual_images, save_grid=save_grid,
|
||||
group_by_prompt=group_by_prompt, save_as_jpg=save_as_jpg, use_GFPGAN=use_GFPGAN,
|
||||
use_RealESRGAN=st.session_state["use_RealESRGAN"] if not loopback else False, loopback=loopback
|
||||
)
|
||||
|
||||
#show a message when the generation is complete.
|
||||
message.success('Render Complete: ' + info + '; Stats: ' + stats, icon="✅")
|
||||
|
||||
except (StopException, KeyError):
|
||||
print(f"Received Streamlit StopException")
|
||||
|
||||
# this will render all the images at the end of the generation but its better if its moved to a second tab inside col2 and shown as a gallery.
|
||||
# use the current col2 first tab to show the preview_img and update it as its generated.
|
||||
#preview_image.image(output_images, width=750)
|
||||
|
||||
#on import run init
|
||||
153
scripts/imglab.py
Normal file
153
scripts/imglab.py
Normal file
@ -0,0 +1,153 @@
|
||||
from webui_streamlit import st, defaults
|
||||
from sd_utils import *
|
||||
|
||||
#home plugin
|
||||
import os
|
||||
from PIL import Image
|
||||
#from bs4 import BeautifulSoup
|
||||
from streamlit.runtime.in_memory_file_manager import in_memory_file_manager
|
||||
from streamlit.elements import image as STImage
|
||||
try:
|
||||
# this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start.
|
||||
from transformers import logging
|
||||
|
||||
logging.set_verbosity_error()
|
||||
except:
|
||||
pass
|
||||
|
||||
class plugin_info():
|
||||
plugname = "imglab"
|
||||
description = "Image Lab"
|
||||
isTab = True
|
||||
displayPriority = 3
|
||||
|
||||
def getLatestGeneratedImagesFromPath():
|
||||
#get the latest images from the generated images folder
|
||||
#get the path to the generated images folder
|
||||
generatedImagesPath = os.path.join(os.getcwd(),'outputs')
|
||||
#get all the files from the folders and subfolders
|
||||
files = []
|
||||
#get the laest 10 images from the output folder without walking the subfolders
|
||||
for r, d, f in os.walk(generatedImagesPath):
|
||||
for file in f:
|
||||
if '.png' in file:
|
||||
files.append(os.path.join(r, file))
|
||||
#sort the files by date
|
||||
files.sort(key=os.path.getmtime)
|
||||
#reverse the list so the latest images are first
|
||||
for f in files:
|
||||
img = Image.open(f)
|
||||
files[files.index(f)] = img
|
||||
#get the latest 10 files
|
||||
#get all the files with the .png or .jpg extension
|
||||
#sort files by date
|
||||
#get the latest 10 files
|
||||
latestFiles = files[-10:]
|
||||
#reverse the list
|
||||
latestFiles.reverse()
|
||||
return latestFiles
|
||||
|
||||
def getImagesFromLexica():
|
||||
#scrape images from lexica.art
|
||||
#get the html from the page
|
||||
#get the html with cookies and javascript
|
||||
apiEndpoint = r'https://lexica.art/api/trpc/prompts.infinitePrompts?batch=1&input=%7B%220%22%3A%7B%22json%22%3A%7B%22limit%22%3A10%2C%22text%22%3A%22%22%2C%22cursor%22%3A10%7D%7D%7D'
|
||||
#REST API call
|
||||
#
|
||||
from requests_html import HTMLSession
|
||||
session = HTMLSession()
|
||||
|
||||
response = session.get(apiEndpoint)
|
||||
#req = requests.Session()
|
||||
#req.headers['user-agent'] = 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/96.0.4664.45 Safari/537.36'
|
||||
#response = req.get(apiEndpoint)
|
||||
print(response.status_code)
|
||||
print(response.text)
|
||||
#get the json from the response
|
||||
#json = response.json()
|
||||
#get the prompts from the json
|
||||
print(response)
|
||||
#session = requests.Session()
|
||||
#parseEndpointJson = session.get(apiEndpoint,headers=headers,verify=False)
|
||||
#print(parseEndpointJson)
|
||||
#print('test2')
|
||||
#page = requests.get("https://lexica.art/", headers={'User-Agent': 'Mozilla/5.0'})
|
||||
#parse the html
|
||||
#soup = BeautifulSoup(page.content, 'html.parser')
|
||||
#find all the images
|
||||
#print(soup)
|
||||
#images = soup.find_all('alt-image')
|
||||
#create a list to store the image urls
|
||||
image_urls = []
|
||||
#loop through the images
|
||||
for image in images:
|
||||
#get the url
|
||||
image_url = image['src']
|
||||
#add it to the list
|
||||
image_urls.append('http://www.lexica.art/'+image_url)
|
||||
#return the list
|
||||
print(image_urls)
|
||||
return image_urls
|
||||
def changeImage():
|
||||
#change the image in the image holder
|
||||
#check if the file is not empty
|
||||
if len(st.session_state['uploaded_file']) > 0:
|
||||
#read the file
|
||||
print('test2')
|
||||
uploaded = st.session_state['uploaded_file'][0].read()
|
||||
#show the image in the image holder
|
||||
st.session_state['previewImg'].empty()
|
||||
st.session_state['previewImg'].image(uploaded,use_column_width=True)
|
||||
def createHTMLGallery(images):
|
||||
html3 = """
|
||||
<div class="gallery-history" style="
|
||||
display: flex;
|
||||
flex-wrap: wrap;
|
||||
align-items: flex-start;">
|
||||
"""
|
||||
mkdwn_array = []
|
||||
for i in images:
|
||||
bImg = i.read()
|
||||
i = Image.save(bImg, 'PNG')
|
||||
width, height = i.size
|
||||
#get random number for the id
|
||||
image_id = "%s" % (str(images.index(i)))
|
||||
(data, mimetype) = STImage._normalize_to_bytes(bImg.getvalue(), width, 'auto')
|
||||
this_file = in_memory_file_manager.add(data, mimetype, image_id)
|
||||
img_str = this_file.url
|
||||
#img_str = 'data:image/png;base64,' + b64encode(image_io.getvalue()).decode('ascii')
|
||||
#get image size
|
||||
|
||||
#make sure the image is not bigger then 150px but keep the aspect ratio
|
||||
if width > 150:
|
||||
height = int(height * (150/width))
|
||||
width = 150
|
||||
if height > 150:
|
||||
width = int(width * (150/height))
|
||||
height = 150
|
||||
|
||||
#mkdwn = f"""<img src="{img_str}" alt="Image" with="200" height="200" />"""
|
||||
mkdwn = f'''<div class="gallery" style="margin: 3px;" >
|
||||
<a href="{img_str}">
|
||||
<img src="{img_str}" alt="Image" width="{width}" height="{height}">
|
||||
</a>
|
||||
</div>
|
||||
'''
|
||||
mkdwn_array.append(mkdwn)
|
||||
html3 += "".join(mkdwn_array)
|
||||
html3 += '</div>'
|
||||
return html3
|
||||
def layout():
|
||||
|
||||
col1, col2 = st.columns(2)
|
||||
with col1:
|
||||
st.session_state['uploaded_file'] = st.file_uploader("Choose an image or images", type=["png", "jpg", "jpeg"],accept_multiple_files=True,on_change=changeImage)
|
||||
if 'previewImg' not in st.session_state:
|
||||
st.session_state['previewImg'] = st.empty()
|
||||
else:
|
||||
if len(st.session_state['uploaded_file']) > 0:
|
||||
st.session_state['previewImg'].empty()
|
||||
st.session_state['previewImg'].image(st.session_state['uploaded_file'][0],use_column_width=True)
|
||||
else:
|
||||
st.session_state['previewImg'] = st.empty()
|
||||
|
||||
1256
scripts/sd_utils.py
Normal file
1256
scripts/sd_utils.py
Normal file
File diff suppressed because it is too large
Load Diff
@ -8,7 +8,7 @@ from diffusers.schedulers import (DDIMScheduler, LMSDiscreteScheduler,
|
||||
PNDMScheduler)
|
||||
from diffusers import ModelMixin
|
||||
|
||||
from .stable_diffusion_pipeline import StableDiffusionPipeline
|
||||
from stable_diffusion_pipeline import StableDiffusionPipeline
|
||||
|
||||
pipeline = StableDiffusionPipeline.from_pretrained(
|
||||
"CompVis/stable-diffusion-v1-4",
|
||||
@ -86,6 +86,8 @@ def walk(
|
||||
use_lerp_for_text=False,
|
||||
scheduler="klms", # choices: default, ddim, klms
|
||||
disable_tqdm=False,
|
||||
upsample=False,
|
||||
fps=30,
|
||||
):
|
||||
"""Generate video frames/a video given a list of prompts and seeds.
|
||||
|
||||
@ -105,10 +107,18 @@ def walk(
|
||||
use_lerp_for_text (bool, optional): Use LERP instead of SLERP for text embeddings when walking. Defaults to False.
|
||||
scheduler (str, optional): Which scheduler to use. Defaults to "klms". Choices are "default", "ddim", "klms".
|
||||
disable_tqdm (bool, optional): Whether to turn off the tqdm progress bars. Defaults to False.
|
||||
upsample (bool, optional): If True, uses Real-ESRGAN to upsample images 4x. Requires it to be installed
|
||||
which you can do by running: `pip install git+https://github.com/xinntao/Real-ESRGAN.git`. Defaults to False.
|
||||
fps (int, optional): The frames per second (fps) that you want the video to use. Does nothing if make_video is False. Defaults to 30.
|
||||
|
||||
Returns:
|
||||
str: Path to video file saved if make_video=True, else None.
|
||||
"""
|
||||
if upsample:
|
||||
from .upsampling import PipelineRealESRGAN
|
||||
|
||||
upsampling_pipeline = PipelineRealESRGAN.from_pretrained('nateraw/real-esrgan')
|
||||
|
||||
pipeline.set_progress_bar_config(disable=disable_tqdm)
|
||||
|
||||
pipeline.scheduler = SCHEDULERS[scheduler]
|
||||
@ -186,8 +196,12 @@ def walk(
|
||||
guidance_scale=guidance_scale,
|
||||
eta=eta,
|
||||
num_inference_steps=num_inference_steps,
|
||||
output_type='pil' if not upsample else 'numpy'
|
||||
)["sample"][0]
|
||||
|
||||
if upsample:
|
||||
im = upsampling_pipeline(im)
|
||||
|
||||
im.save(output_path / ("frame%06d.jpg" % frame_index))
|
||||
frame_index += 1
|
||||
|
||||
@ -195,7 +209,7 @@ def walk(
|
||||
latents_a = latents_b
|
||||
|
||||
if make_video:
|
||||
return make_video_ffmpeg(output_path, f"{name}.mp4")
|
||||
return make_video_ffmpeg(output_path, f"{name}.mp4", fps=fps)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
332
scripts/txt2img.py
Normal file
332
scripts/txt2img.py
Normal file
@ -0,0 +1,332 @@
|
||||
from webui_streamlit import st
|
||||
from sd_utils import *
|
||||
|
||||
from streamlit import StopException
|
||||
|
||||
import os
|
||||
from typing import Union
|
||||
from io import BytesIO
|
||||
from ldm.models.diffusion.ddim import DDIMSampler
|
||||
from ldm.models.diffusion.plms import PLMSSampler
|
||||
|
||||
from streamlit.runtime.in_memory_file_manager import in_memory_file_manager
|
||||
from streamlit.elements import image as STImage
|
||||
try:
|
||||
# this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start.
|
||||
from transformers import logging
|
||||
|
||||
logging.set_verbosity_error()
|
||||
except:
|
||||
pass
|
||||
|
||||
class plugin_info():
|
||||
plugname = "txt2img"
|
||||
description = "Text to Image"
|
||||
isTab = True
|
||||
displayPriority = 1
|
||||
|
||||
|
||||
if os.path.exists(os.path.join(defaults.general.GFPGAN_dir, "experiments", "pretrained_models", "GFPGANv1.3.pth")):
|
||||
GFPGAN_available = True
|
||||
else:
|
||||
GFPGAN_available = False
|
||||
|
||||
if os.path.exists(os.path.join(defaults.general.RealESRGAN_dir, "experiments","pretrained_models", f"{defaults.general.RealESRGAN_model}.pth")):
|
||||
RealESRGAN_available = True
|
||||
else:
|
||||
RealESRGAN_available = False
|
||||
|
||||
#
|
||||
def txt2img(prompt: str, ddim_steps: int, sampler_name: str, realesrgan_model_name: str,
|
||||
n_iter: int, batch_size: int, cfg_scale: float, seed: Union[int, str, None],
|
||||
height: int, width: int, separate_prompts:bool = False, normalize_prompt_weights:bool = True,
|
||||
save_individual_images: bool = True, save_grid: bool = True, group_by_prompt: bool = True,
|
||||
save_as_jpg: bool = True, use_GFPGAN: bool = True, use_RealESRGAN: bool = True,
|
||||
RealESRGAN_model: str = "RealESRGAN_x4plus_anime_6B", fp = None, variant_amount: float = None,
|
||||
variant_seed: int = None, ddim_eta:float = 0.0, write_info_files:bool = True):
|
||||
|
||||
outpath = defaults.general.outdir_txt2img or defaults.general.outdir or "outputs/txt2img-samples"
|
||||
|
||||
seed = seed_to_int(seed)
|
||||
|
||||
#prompt_matrix = 0 in toggles
|
||||
#normalize_prompt_weights = 1 in toggles
|
||||
#skip_save = 2 not in toggles
|
||||
#save_grid = 3 not in toggles
|
||||
#sort_samples = 4 in toggles
|
||||
#write_info_files = 5 in toggles
|
||||
#jpg_sample = 6 in toggles
|
||||
#use_GFPGAN = 7 in toggles
|
||||
#use_RealESRGAN = 8 in toggles
|
||||
|
||||
if sampler_name == 'PLMS':
|
||||
sampler = PLMSSampler(st.session_state["model"])
|
||||
elif sampler_name == 'DDIM':
|
||||
sampler = DDIMSampler(st.session_state["model"])
|
||||
elif sampler_name == 'k_dpm_2_a':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'dpm_2_ancestral')
|
||||
elif sampler_name == 'k_dpm_2':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'dpm_2')
|
||||
elif sampler_name == 'k_euler_a':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'euler_ancestral')
|
||||
elif sampler_name == 'k_euler':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'euler')
|
||||
elif sampler_name == 'k_heun':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'heun')
|
||||
elif sampler_name == 'k_lms':
|
||||
sampler = KDiffusionSampler(st.session_state["model"],'lms')
|
||||
else:
|
||||
raise Exception("Unknown sampler: " + sampler_name)
|
||||
|
||||
def init():
|
||||
pass
|
||||
|
||||
def sample(init_data, x, conditioning, unconditional_conditioning, sampler_name):
|
||||
samples_ddim, _ = sampler.sample(S=ddim_steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=cfg_scale,
|
||||
unconditional_conditioning=unconditional_conditioning, eta=ddim_eta, x_T=x, img_callback=generation_callback,
|
||||
log_every_t=int(defaults.general.update_preview_frequency))
|
||||
|
||||
return samples_ddim
|
||||
|
||||
#try:
|
||||
output_images, seed, info, stats = process_images(
|
||||
outpath=outpath,
|
||||
func_init=init,
|
||||
func_sample=sample,
|
||||
prompt=prompt,
|
||||
seed=seed,
|
||||
sampler_name=sampler_name,
|
||||
save_grid=save_grid,
|
||||
batch_size=batch_size,
|
||||
n_iter=n_iter,
|
||||
steps=ddim_steps,
|
||||
cfg_scale=cfg_scale,
|
||||
width=width,
|
||||
height=height,
|
||||
prompt_matrix=separate_prompts,
|
||||
use_GFPGAN=use_GFPGAN,
|
||||
use_RealESRGAN=use_RealESRGAN,
|
||||
realesrgan_model_name=realesrgan_model_name,
|
||||
fp=fp,
|
||||
ddim_eta=ddim_eta,
|
||||
normalize_prompt_weights=normalize_prompt_weights,
|
||||
save_individual_images=save_individual_images,
|
||||
sort_samples=group_by_prompt,
|
||||
write_info_files=write_info_files,
|
||||
jpg_sample=save_as_jpg,
|
||||
variant_amount=variant_amount,
|
||||
variant_seed=variant_seed,
|
||||
)
|
||||
|
||||
del sampler
|
||||
|
||||
return output_images, seed, info, stats
|
||||
|
||||
#except RuntimeError as e:
|
||||
#err = e
|
||||
#err_msg = f'CRASHED:<br><textarea rows="5" style="color:white;background: black;width: -webkit-fill-available;font-family: monospace;font-size: small;font-weight: bold;">{str(e)}</textarea><br><br>Please wait while the program restarts.'
|
||||
#stats = err_msg
|
||||
#return [], seed, 'err', stats
|
||||
|
||||
def layout():
|
||||
with st.form("txt2img-inputs"):
|
||||
st.session_state["generation_mode"] = "txt2img"
|
||||
|
||||
input_col1, generate_col1 = st.columns([10,1])
|
||||
|
||||
with input_col1:
|
||||
#prompt = st.text_area("Input Text","")
|
||||
prompt = st.text_input("Input Text","", placeholder="A corgi wearing a top hat as an oil painting.")
|
||||
|
||||
# Every form must have a submit button, the extra blank spaces is a temp way to align it with the input field. Needs to be done in CSS or some other way.
|
||||
generate_col1.write("")
|
||||
generate_col1.write("")
|
||||
generate_button = generate_col1.form_submit_button("Generate")
|
||||
|
||||
# creating the page layout using columns
|
||||
col1, col2, col3 = st.columns([1,2,1], gap="large")
|
||||
|
||||
with col1:
|
||||
width = st.slider("Width:", min_value=64, max_value=1024, value=defaults.txt2img.width, step=64)
|
||||
height = st.slider("Height:", min_value=64, max_value=1024, value=defaults.txt2img.height, step=64)
|
||||
cfg_scale = st.slider("CFG (Classifier Free Guidance Scale):", min_value=1.0, max_value=30.0, value=defaults.txt2img.cfg_scale, step=0.5, help="How strongly the image should follow the prompt.")
|
||||
seed = st.text_input("Seed:", value=defaults.txt2img.seed, help=" The seed to use, if left blank a random seed will be generated.")
|
||||
batch_count = st.slider("Batch count.", min_value=1, max_value=100, value=defaults.txt2img.batch_count, step=1, help="How many iterations or batches of images to generate in total.")
|
||||
#batch_size = st.slider("Batch size", min_value=1, max_value=250, value=defaults.txt2img.batch_size, step=1,
|
||||
#help="How many images are at once in a batch.\
|
||||
#It increases the VRAM usage a lot but if you have enough VRAM it can reduce the time it takes to finish generation as more images are generated at once.\
|
||||
#Default: 1")
|
||||
|
||||
with st.expander("Preview Settings"):
|
||||
st.session_state["update_preview"] = st.checkbox("Update Image Preview", value=defaults.txt2img.update_preview,
|
||||
help="If enabled the image preview will be updated during the generation instead of at the end. \
|
||||
You can use the Update Preview \Frequency option bellow to customize how frequent it's updated. \
|
||||
By default this is enabled and the frequency is set to 1 step.")
|
||||
|
||||
st.session_state["update_preview_frequency"] = st.text_input("Update Image Preview Frequency", value=defaults.txt2img.update_preview_frequency,
|
||||
help="Frequency in steps at which the the preview image is updated. By default the frequency \
|
||||
is set to 1 step.")
|
||||
|
||||
with col2:
|
||||
preview_tab, gallery_tab = st.tabs(["Preview", "Gallery"])
|
||||
|
||||
with preview_tab:
|
||||
#st.write("Image")
|
||||
#Image for testing
|
||||
#image = Image.open(requests.get("https://icon-library.com/images/image-placeholder-icon/image-placeholder-icon-13.jpg", stream=True).raw).convert('RGB')
|
||||
#new_image = image.resize((175, 240))
|
||||
#preview_image = st.image(image)
|
||||
|
||||
# create an empty container for the image, progress bar, etc so we can update it later and use session_state to hold them globally.
|
||||
st.session_state["preview_image"] = st.empty()
|
||||
|
||||
st.session_state["loading"] = st.empty()
|
||||
|
||||
st.session_state["progress_bar_text"] = st.empty()
|
||||
st.session_state["progress_bar"] = st.empty()
|
||||
|
||||
message = st.empty()
|
||||
|
||||
with col3:
|
||||
st.session_state.sampling_steps = st.slider("Sampling Steps", value=defaults.txt2img.sampling_steps, min_value=1, max_value=250)
|
||||
|
||||
sampler_name_list = ["k_lms", "k_euler", "k_euler_a", "k_dpm_2", "k_dpm_2_a", "k_heun", "PLMS", "DDIM"]
|
||||
sampler_name = st.selectbox("Sampling method", sampler_name_list,
|
||||
index=sampler_name_list.index(defaults.txt2img.default_sampler), help="Sampling method to use. Default: k_euler")
|
||||
|
||||
|
||||
|
||||
#basic_tab, advanced_tab = st.tabs(["Basic", "Advanced"])
|
||||
|
||||
#with basic_tab:
|
||||
#summit_on_enter = st.radio("Submit on enter?", ("Yes", "No"), horizontal=True,
|
||||
#help="Press the Enter key to summit, when 'No' is selected you can use the Enter key to write multiple lines.")
|
||||
|
||||
with st.expander("Advanced"):
|
||||
separate_prompts = st.checkbox("Create Prompt Matrix.", value=False, help="Separate multiple prompts using the `|` character, and get all combinations of them.")
|
||||
normalize_prompt_weights = st.checkbox("Normalize Prompt Weights.", value=True, help="Ensure the sum of all weights add up to 1.0")
|
||||
save_individual_images = st.checkbox("Save individual images.", value=True, help="Save each image generated before any filter or enhancement is applied.")
|
||||
save_grid = st.checkbox("Save grid",value=True, help="Save a grid with all the images generated into a single image.")
|
||||
group_by_prompt = st.checkbox("Group results by prompt", value=True,
|
||||
help="Saves all the images with the same prompt into the same folder. When using a prompt matrix each prompt combination will have its own folder.")
|
||||
write_info_files = st.checkbox("Write Info file", value=True, help="Save a file next to the image with informartion about the generation.")
|
||||
save_as_jpg = st.checkbox("Save samples as jpg", value=False, help="Saves the images as jpg instead of png.")
|
||||
|
||||
if GFPGAN_available:
|
||||
use_GFPGAN = st.checkbox("Use GFPGAN", value=defaults.txt2img.use_GFPGAN, help="Uses the GFPGAN model to improve faces after the generation. This greatly improve the quality and consistency of faces but uses extra VRAM. Disable if you need the extra VRAM.")
|
||||
else:
|
||||
use_GFPGAN = False
|
||||
|
||||
if RealESRGAN_available:
|
||||
use_RealESRGAN = st.checkbox("Use RealESRGAN", value=defaults.txt2img.use_RealESRGAN, help="Uses the RealESRGAN model to upscale the images after the generation. This greatly improve the quality and lets you have high resolution images but uses extra VRAM. Disable if you need the extra VRAM.")
|
||||
RealESRGAN_model = st.selectbox("RealESRGAN model", ["RealESRGAN_x4plus", "RealESRGAN_x4plus_anime_6B"], index=0)
|
||||
else:
|
||||
use_RealESRGAN = False
|
||||
RealESRGAN_model = "RealESRGAN_x4plus"
|
||||
|
||||
variant_amount = st.slider("Variant Amount:", value=defaults.txt2img.variant_amount, min_value=0.0, max_value=1.0, step=0.01)
|
||||
variant_seed = st.text_input("Variant Seed:", value=defaults.txt2img.seed, help="The seed to use when generating a variant, if left blank a random seed will be generated.")
|
||||
galleryCont = st.empty()
|
||||
|
||||
if generate_button:
|
||||
#print("Loading models")
|
||||
# load the models when we hit the generate button for the first time, it wont be loaded after that so dont worry.
|
||||
load_models(False, use_GFPGAN, use_RealESRGAN, RealESRGAN_model)
|
||||
|
||||
try:
|
||||
output_images, seeds, info, stats = txt2img(prompt, st.session_state.sampling_steps, sampler_name, RealESRGAN_model, batch_count, 1,
|
||||
cfg_scale, seed, height, width, separate_prompts, normalize_prompt_weights, save_individual_images,
|
||||
save_grid, group_by_prompt, save_as_jpg, use_GFPGAN, use_RealESRGAN, RealESRGAN_model, fp=defaults.general.fp,
|
||||
variant_amount=variant_amount, variant_seed=variant_seed, write_info_files=write_info_files)
|
||||
|
||||
message.success('Done!', 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()
|
||||
with col1_cont:
|
||||
with col1:
|
||||
st.image(st.session_state['latestImages'][0])
|
||||
st.image(st.session_state['latestImages'][3])
|
||||
st.image(st.session_state['latestImages'][6])
|
||||
with col2_cont:
|
||||
with col2:
|
||||
st.image(st.session_state['latestImages'][1])
|
||||
st.image(st.session_state['latestImages'][4])
|
||||
st.image(st.session_state['latestImages'][7])
|
||||
with col3_cont:
|
||||
with col3:
|
||||
st.image(st.session_state['latestImages'][2])
|
||||
st.image(st.session_state['latestImages'][5])
|
||||
st.image(st.session_state['latestImages'][8])
|
||||
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]
|
||||
except (StopException, KeyError):
|
||||
print(f"Received Streamlit StopException")
|
||||
|
||||
# this will render all the images at the end of the generation but its better if its moved to a second tab inside col2 and shown as a gallery.
|
||||
# use the current col2 first tab to show the preview_img and update it as its generated.
|
||||
#preview_image.image(output_images)
|
||||
|
||||
#on import run init
|
||||
def createHTMLGallery(images,info):
|
||||
html3 = """
|
||||
<div class="gallery-history" style="
|
||||
display: flex;
|
||||
flex-wrap: wrap;
|
||||
align-items: flex-start;">
|
||||
"""
|
||||
mkdwn_array = []
|
||||
for i in images:
|
||||
try:
|
||||
seed = info[images.index(i)]
|
||||
except:
|
||||
seed = ' '
|
||||
image_io = BytesIO()
|
||||
i.save(image_io, 'PNG')
|
||||
width, height = i.size
|
||||
#get random number for the id
|
||||
image_id = "%s" % (str(images.index(i)))
|
||||
(data, mimetype) = STImage._normalize_to_bytes(image_io.getvalue(), width, 'auto')
|
||||
this_file = in_memory_file_manager.add(data, mimetype, image_id)
|
||||
img_str = this_file.url
|
||||
#img_str = 'data:image/png;base64,' + b64encode(image_io.getvalue()).decode('ascii')
|
||||
#get image size
|
||||
|
||||
#make sure the image is not bigger then 150px but keep the aspect ratio
|
||||
if width > 150:
|
||||
height = int(height * (150/width))
|
||||
width = 150
|
||||
if height > 150:
|
||||
width = int(width * (150/height))
|
||||
height = 150
|
||||
|
||||
#mkdwn = f"""<img src="{img_str}" alt="Image" with="200" height="200" />"""
|
||||
mkdwn = f'''<div class="gallery" style="margin: 3px;" >
|
||||
<a href="{img_str}">
|
||||
<img src="{img_str}" alt="Image" width="{width}" height="{height}">
|
||||
</a>
|
||||
<div class="desc" style="text-align: center; opacity: 40%;">{seed}</div>
|
||||
</div>
|
||||
'''
|
||||
mkdwn_array.append(mkdwn)
|
||||
html3 += "".join(mkdwn_array)
|
||||
html3 += '</div>'
|
||||
return html3
|
||||
678
scripts/txt2vid.py
Normal file
678
scripts/txt2vid.py
Normal file
@ -0,0 +1,678 @@
|
||||
from webui_streamlit import st
|
||||
from sd_utils import *
|
||||
|
||||
from streamlit import StopException
|
||||
|
||||
import os
|
||||
from PIL import Image
|
||||
import torch
|
||||
import numpy as np
|
||||
import time
|
||||
import torch
|
||||
from torch import autocast
|
||||
from io import BytesIO
|
||||
# we use python-slugify to make the filenames safe for windows and linux, its better than doing it manually
|
||||
# install it with 'pip install python-slugify'
|
||||
from slugify import slugify
|
||||
from streamlit.runtime.in_memory_file_manager import in_memory_file_manager
|
||||
from streamlit.elements import image as STImage
|
||||
|
||||
try:
|
||||
# this silences the annoying "Some weights of the model checkpoint were not used when initializing..." message at start.
|
||||
from transformers import logging
|
||||
|
||||
logging.set_verbosity_error()
|
||||
except:
|
||||
pass
|
||||
|
||||
class plugin_info():
|
||||
plugname = "txt2img"
|
||||
description = "Text to Image"
|
||||
isTab = True
|
||||
displayPriority = 1
|
||||
|
||||
|
||||
if os.path.exists(os.path.join(defaults.general.GFPGAN_dir, "experiments", "pretrained_models", "GFPGANv1.3.pth")):
|
||||
GFPGAN_available = True
|
||||
else:
|
||||
GFPGAN_available = False
|
||||
|
||||
if os.path.exists(os.path.join(defaults.general.RealESRGAN_dir, "experiments","pretrained_models", f"{defaults.general.RealESRGAN_model}.pth")):
|
||||
RealESRGAN_available = True
|
||||
else:
|
||||
RealESRGAN_available = False
|
||||
|
||||
#
|
||||
# -----------------------------------------------------------------------------
|
||||
|
||||
@torch.no_grad()
|
||||
def diffuse(
|
||||
pipe,
|
||||
cond_embeddings, # text conditioning, should be (1, 77, 768)
|
||||
cond_latents, # image conditioning, should be (1, 4, 64, 64)
|
||||
num_inference_steps,
|
||||
cfg_scale,
|
||||
eta,
|
||||
):
|
||||
|
||||
torch_device = cond_latents.get_device()
|
||||
|
||||
# classifier guidance: add the unconditional embedding
|
||||
max_length = cond_embeddings.shape[1] # 77
|
||||
uncond_input = pipe.tokenizer([""], padding="max_length", max_length=max_length, return_tensors="pt")
|
||||
uncond_embeddings = pipe.text_encoder(uncond_input.input_ids.to(torch_device))[0]
|
||||
text_embeddings = torch.cat([uncond_embeddings, cond_embeddings])
|
||||
|
||||
# if we use LMSDiscreteScheduler, let's make sure latents are mulitplied by sigmas
|
||||
if isinstance(pipe.scheduler, LMSDiscreteScheduler):
|
||||
cond_latents = cond_latents * pipe.scheduler.sigmas[0]
|
||||
|
||||
# init the scheduler
|
||||
accepts_offset = "offset" in set(inspect.signature(pipe.scheduler.set_timesteps).parameters.keys())
|
||||
extra_set_kwargs = {}
|
||||
if accepts_offset:
|
||||
extra_set_kwargs["offset"] = 1
|
||||
|
||||
pipe.scheduler.set_timesteps(num_inference_steps + st.session_state.sampling_steps, **extra_set_kwargs)
|
||||
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
||||
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
||||
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
||||
# and should be between [0, 1]
|
||||
accepts_eta = "eta" in set(inspect.signature(pipe.scheduler.step).parameters.keys())
|
||||
extra_step_kwargs = {}
|
||||
if accepts_eta:
|
||||
extra_step_kwargs["eta"] = eta
|
||||
|
||||
|
||||
step_counter = 0
|
||||
inference_counter = 0
|
||||
current_chunk_speed = 0
|
||||
previous_chunk_speed = 0
|
||||
|
||||
# diffuse!
|
||||
for i, t in enumerate(pipe.scheduler.timesteps):
|
||||
start = timeit.default_timer()
|
||||
|
||||
#status_text.text(f"Running step: {step_counter}{total_number_steps} {percent} | {duration:.2f}{speed}")
|
||||
|
||||
# expand the latents for classifier free guidance
|
||||
latent_model_input = torch.cat([cond_latents] * 2)
|
||||
if isinstance(pipe.scheduler, LMSDiscreteScheduler):
|
||||
sigma = pipe.scheduler.sigmas[i]
|
||||
latent_model_input = latent_model_input / ((sigma**2 + 1) ** 0.5)
|
||||
|
||||
# predict the noise residual
|
||||
noise_pred = pipe.unet(latent_model_input, t, encoder_hidden_states=text_embeddings)["sample"]
|
||||
|
||||
# cfg
|
||||
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
||||
noise_pred = noise_pred_uncond + cfg_scale * (noise_pred_text - noise_pred_uncond)
|
||||
|
||||
# compute the previous noisy sample x_t -> x_t-1
|
||||
if isinstance(pipe.scheduler, LMSDiscreteScheduler):
|
||||
cond_latents = pipe.scheduler.step(noise_pred, i, cond_latents, **extra_step_kwargs)["prev_sample"]
|
||||
else:
|
||||
cond_latents = pipe.scheduler.step(noise_pred, t, cond_latents, **extra_step_kwargs)["prev_sample"]
|
||||
|
||||
#print (st.session_state["update_preview_frequency"])
|
||||
#update the preview image if it is enabled and the frequency matches the step_counter
|
||||
if defaults.general.update_preview:
|
||||
step_counter += 1
|
||||
|
||||
if st.session_state.dynamic_preview_frequency:
|
||||
current_chunk_speed, previous_chunk_speed, defaults.general.update_preview_frequency = optimize_update_preview_frequency(
|
||||
current_chunk_speed, previous_chunk_speed, defaults.general.update_preview_frequency)
|
||||
|
||||
if defaults.general.update_preview_frequency == step_counter or step_counter == st.session_state.sampling_steps:
|
||||
#scale and decode the image latents with vae
|
||||
cond_latents_2 = 1 / 0.18215 * cond_latents
|
||||
image_2 = pipe.vae.decode(cond_latents_2)
|
||||
|
||||
# generate output numpy image as uint8
|
||||
image_2 = (image_2 / 2 + 0.5).clamp(0, 1)
|
||||
image_2 = image_2.cpu().permute(0, 2, 3, 1).numpy()
|
||||
image_2 = (image_2[0] * 255).astype(np.uint8)
|
||||
|
||||
st.session_state["preview_image"].image(image_2)
|
||||
|
||||
step_counter = 0
|
||||
|
||||
duration = timeit.default_timer() - start
|
||||
|
||||
current_chunk_speed = duration
|
||||
|
||||
if duration >= 1:
|
||||
speed = "s/it"
|
||||
else:
|
||||
speed = "it/s"
|
||||
duration = 1 / duration
|
||||
|
||||
if i > st.session_state.sampling_steps:
|
||||
inference_counter += 1
|
||||
inference_percent = int(100 * float(inference_counter if inference_counter < num_inference_steps else num_inference_steps)/float(num_inference_steps))
|
||||
inference_progress = f"{inference_counter if inference_counter < num_inference_steps else num_inference_steps}/{num_inference_steps} {inference_percent}% "
|
||||
else:
|
||||
inference_progress = ""
|
||||
|
||||
percent = int(100 * float(i+1 if i+1 < st.session_state.sampling_steps else st.session_state.sampling_steps)/float(st.session_state.sampling_steps))
|
||||
frames_percent = int(100 * float(st.session_state.current_frame if st.session_state.current_frame < st.session_state.max_frames else st.session_state.max_frames)/float(st.session_state.max_frames))
|
||||
|
||||
st.session_state["progress_bar_text"].text(
|
||||
f"Running step: {i+1 if i+1 < st.session_state.sampling_steps else st.session_state.sampling_steps}/{st.session_state.sampling_steps} "
|
||||
f"{percent if percent < 100 else 100}% {inference_progress}{duration:.2f}{speed} | "
|
||||
f"Frame: {st.session_state.current_frame if st.session_state.current_frame < st.session_state.max_frames else st.session_state.max_frames}/{st.session_state.max_frames} "
|
||||
f"{frames_percent if frames_percent < 100 else 100}% {st.session_state.frame_duration:.2f}{st.session_state.frame_speed}"
|
||||
)
|
||||
st.session_state["progress_bar"].progress(percent if percent < 100 else 100)
|
||||
|
||||
# scale and decode the image latents with vae
|
||||
cond_latents = 1 / 0.18215 * cond_latents
|
||||
image = pipe.vae.decode(cond_latents)
|
||||
|
||||
# generate output numpy image as uint8
|
||||
image = (image / 2 + 0.5).clamp(0, 1)
|
||||
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
||||
image = (image[0] * 255).astype(np.uint8)
|
||||
|
||||
return image
|
||||
|
||||
#
|
||||
def txt2vid(
|
||||
# --------------------------------------
|
||||
# args you probably want to change
|
||||
prompts = ["blueberry spaghetti", "strawberry spaghetti"], # prompt to dream about
|
||||
gpu:int = defaults.general.gpu, # id of the gpu to run on
|
||||
#name:str = 'test', # name of this project, for the output directory
|
||||
#rootdir:str = defaults.general.outdir,
|
||||
num_steps:int = 200, # number of steps between each pair of sampled points
|
||||
max_frames:int = 10000, # number of frames to write and then exit the script
|
||||
num_inference_steps:int = 50, # more (e.g. 100, 200 etc) can create slightly better images
|
||||
cfg_scale:float = 5.0, # can depend on the prompt. usually somewhere between 3-10 is good
|
||||
do_loop = False,
|
||||
use_lerp_for_text = False,
|
||||
seeds = None,
|
||||
quality:int = 100, # for jpeg compression of the output images
|
||||
eta:float = 0.0,
|
||||
width:int = 256,
|
||||
height:int = 256,
|
||||
weights_path = "CompVis/stable-diffusion-v1-4",
|
||||
scheduler="klms", # choices: default, ddim, klms
|
||||
disable_tqdm = False,
|
||||
#-----------------------------------------------
|
||||
beta_start = 0.0001,
|
||||
beta_end = 0.00012,
|
||||
beta_schedule = "scaled_linear"
|
||||
):
|
||||
"""
|
||||
prompt = ["blueberry spaghetti", "strawberry spaghetti"], # prompt to dream about
|
||||
gpu:int = defaults.general.gpu, # id of the gpu to run on
|
||||
#name:str = 'test', # name of this project, for the output directory
|
||||
#rootdir:str = defaults.general.outdir,
|
||||
num_steps:int = 200, # number of steps between each pair of sampled points
|
||||
max_frames:int = 10000, # number of frames to write and then exit the script
|
||||
num_inference_steps:int = 50, # more (e.g. 100, 200 etc) can create slightly better images
|
||||
cfg_scale:float = 5.0, # can depend on the prompt. usually somewhere between 3-10 is good
|
||||
do_loop = False,
|
||||
use_lerp_for_text = False,
|
||||
seed = None,
|
||||
quality:int = 100, # for jpeg compression of the output images
|
||||
eta:float = 0.0,
|
||||
width:int = 256,
|
||||
height:int = 256,
|
||||
weights_path = "CompVis/stable-diffusion-v1-4",
|
||||
scheduler="klms", # choices: default, ddim, klms
|
||||
disable_tqdm = False,
|
||||
beta_start = 0.0001,
|
||||
beta_end = 0.00012,
|
||||
beta_schedule = "scaled_linear"
|
||||
"""
|
||||
mem_mon = MemUsageMonitor('MemMon')
|
||||
mem_mon.start()
|
||||
|
||||
|
||||
seeds = seed_to_int(seeds)
|
||||
|
||||
# We add an extra frame because most
|
||||
# of the time the first frame is just the noise.
|
||||
max_frames +=1
|
||||
|
||||
assert torch.cuda.is_available()
|
||||
assert height % 8 == 0 and width % 8 == 0
|
||||
torch.manual_seed(seeds)
|
||||
torch_device = f"cuda:{gpu}"
|
||||
|
||||
# init the output dir
|
||||
sanitized_prompt = slugify(prompts)
|
||||
|
||||
full_path = os.path.join(os.getcwd(), defaults.general.outdir, "txt2vid-samples", "samples", sanitized_prompt)
|
||||
|
||||
if len(full_path) > 220:
|
||||
sanitized_prompt = sanitized_prompt[:220-len(full_path)]
|
||||
full_path = os.path.join(os.getcwd(), defaults.general.outdir, "txt2vid-samples", "samples", sanitized_prompt)
|
||||
|
||||
os.makedirs(full_path, exist_ok=True)
|
||||
|
||||
# Write prompt info to file in output dir so we can keep track of what we did
|
||||
if st.session_state.write_info_files:
|
||||
with open(os.path.join(full_path , f'{slugify(str(seeds))}_config.json' if len(prompts) > 1 else "prompts_config.json"), "w") as outfile:
|
||||
outfile.write(json.dumps(
|
||||
dict(
|
||||
prompts = prompts,
|
||||
gpu = gpu,
|
||||
num_steps = num_steps,
|
||||
max_frames = max_frames,
|
||||
num_inference_steps = num_inference_steps,
|
||||
cfg_scale = cfg_scale,
|
||||
do_loop = do_loop,
|
||||
use_lerp_for_text = use_lerp_for_text,
|
||||
seeds = seeds,
|
||||
quality = quality,
|
||||
eta = eta,
|
||||
width = width,
|
||||
height = height,
|
||||
weights_path = weights_path,
|
||||
scheduler=scheduler,
|
||||
disable_tqdm = disable_tqdm,
|
||||
beta_start = beta_start,
|
||||
beta_end = beta_end,
|
||||
beta_schedule = beta_schedule
|
||||
),
|
||||
indent=2,
|
||||
sort_keys=False,
|
||||
))
|
||||
|
||||
#print(scheduler)
|
||||
default_scheduler = PNDMScheduler(
|
||||
beta_start=beta_start, beta_end=beta_end, beta_schedule=beta_schedule
|
||||
)
|
||||
# ------------------------------------------------------------------------------
|
||||
#Schedulers
|
||||
ddim_scheduler = DDIMScheduler(
|
||||
beta_start=beta_start,
|
||||
beta_end=beta_end,
|
||||
beta_schedule=beta_schedule,
|
||||
clip_sample=False,
|
||||
set_alpha_to_one=False,
|
||||
)
|
||||
|
||||
klms_scheduler = LMSDiscreteScheduler(
|
||||
beta_start=beta_start, beta_end=beta_end, beta_schedule=beta_schedule
|
||||
)
|
||||
|
||||
SCHEDULERS = dict(default=default_scheduler, ddim=ddim_scheduler, klms=klms_scheduler)
|
||||
|
||||
# ------------------------------------------------------------------------------
|
||||
|
||||
#if weights_path == "Stable Diffusion v1.4":
|
||||
#weights_path = "CompVis/stable-diffusion-v1-4"
|
||||
#else:
|
||||
#weights_path = os.path.join("./models", "custom", f"{weights_path}.ckpt")
|
||||
|
||||
try:
|
||||
if "model" in st.session_state:
|
||||
del st.session_state["model"]
|
||||
except:
|
||||
pass
|
||||
|
||||
#print (st.session_state["weights_path"] != weights_path)
|
||||
|
||||
try:
|
||||
if not st.session_state["pipe"] or st.session_state["weights_path"] != weights_path:
|
||||
if st.session_state["weights_path"] != weights_path:
|
||||
del st.session_state["weights_path"]
|
||||
|
||||
st.session_state["weights_path"] = weights_path
|
||||
st.session_state["pipe"] = StableDiffusionPipeline.from_pretrained(
|
||||
weights_path,
|
||||
use_local_file=True,
|
||||
use_auth_token=True,
|
||||
#torch_dtype=torch.float16 if not defaults.general.no_half else None,
|
||||
revision="fp16" if not defaults.general.no_half else None
|
||||
)
|
||||
|
||||
st.session_state["pipe"].unet.to(torch_device)
|
||||
st.session_state["pipe"].vae.to(torch_device)
|
||||
st.session_state["pipe"].text_encoder.to(torch_device)
|
||||
print("Tx2Vid Model Loaded")
|
||||
else:
|
||||
print("Tx2Vid Model already Loaded")
|
||||
|
||||
except:
|
||||
#del st.session_state["weights_path"]
|
||||
#del st.session_state["pipe"]
|
||||
|
||||
st.session_state["weights_path"] = weights_path
|
||||
st.session_state["pipe"] = StableDiffusionPipeline.from_pretrained(
|
||||
weights_path,
|
||||
use_local_file=True,
|
||||
use_auth_token=True,
|
||||
#torch_dtype=torch.float16 if not defaults.general.no_half else None,
|
||||
revision="fp16" if not defaults.general.no_half else None
|
||||
)
|
||||
|
||||
st.session_state["pipe"].unet.to(torch_device)
|
||||
st.session_state["pipe"].vae.to(torch_device)
|
||||
st.session_state["pipe"].text_encoder.to(torch_device)
|
||||
print("Tx2Vid Model Loaded")
|
||||
|
||||
st.session_state["pipe"].scheduler = SCHEDULERS[scheduler]
|
||||
|
||||
# get the conditional text embeddings based on the prompt
|
||||
text_input = st.session_state["pipe"].tokenizer(prompts, padding="max_length", max_length=st.session_state["pipe"].tokenizer.model_max_length, truncation=True, return_tensors="pt")
|
||||
cond_embeddings = st.session_state["pipe"].text_encoder(text_input.input_ids.to(torch_device))[0] # shape [1, 77, 768]
|
||||
|
||||
# sample a source
|
||||
init1 = torch.randn((1, st.session_state["pipe"].unet.in_channels, height // 8, width // 8), device=torch_device)
|
||||
|
||||
if do_loop:
|
||||
prompts = [prompts, prompts]
|
||||
seeds = [seeds, seeds]
|
||||
#first_seed, *seeds = seeds
|
||||
#prompts.append(prompts)
|
||||
#seeds.append(first_seed)
|
||||
|
||||
|
||||
# iterate the loop
|
||||
frames = []
|
||||
frame_index = 0
|
||||
|
||||
st.session_state["frame_total_duration"] = 0
|
||||
st.session_state["frame_total_speed"] = 0
|
||||
|
||||
try:
|
||||
while frame_index < max_frames:
|
||||
st.session_state["frame_duration"] = 0
|
||||
st.session_state["frame_speed"] = 0
|
||||
st.session_state["current_frame"] = frame_index
|
||||
|
||||
# sample the destination
|
||||
init2 = torch.randn((1, st.session_state["pipe"].unet.in_channels, height // 8, width // 8), device=torch_device)
|
||||
|
||||
for i, t in enumerate(np.linspace(0, 1, num_steps)):
|
||||
start = timeit.default_timer()
|
||||
print(f"COUNT: {frame_index+1}/{num_steps}")
|
||||
|
||||
#if use_lerp_for_text:
|
||||
#init = torch.lerp(init1, init2, float(t))
|
||||
#else:
|
||||
#init = slerp(gpu, float(t), init1, init2)
|
||||
|
||||
init = slerp(gpu, float(t), init1, init2)
|
||||
|
||||
with autocast("cuda"):
|
||||
image = diffuse(st.session_state["pipe"], cond_embeddings, init, num_inference_steps, cfg_scale, eta)
|
||||
|
||||
im = Image.fromarray(image)
|
||||
outpath = os.path.join(full_path, 'frame%06d.png' % frame_index)
|
||||
im.save(outpath, quality=quality)
|
||||
|
||||
# send the image to the UI to update it
|
||||
#st.session_state["preview_image"].image(im)
|
||||
|
||||
#append the frames to the frames list so we can use them later.
|
||||
frames.append(np.asarray(im))
|
||||
|
||||
#increase frame_index counter.
|
||||
frame_index += 1
|
||||
|
||||
st.session_state["current_frame"] = frame_index
|
||||
|
||||
duration = timeit.default_timer() - start
|
||||
|
||||
if duration >= 1:
|
||||
speed = "s/it"
|
||||
else:
|
||||
speed = "it/s"
|
||||
duration = 1 / duration
|
||||
|
||||
st.session_state["frame_duration"] = duration
|
||||
st.session_state["frame_speed"] = speed
|
||||
|
||||
init1 = init2
|
||||
|
||||
except StopException:
|
||||
pass
|
||||
|
||||
|
||||
if st.session_state['save_video']:
|
||||
# write video to memory
|
||||
#output = io.BytesIO()
|
||||
#writer = imageio.get_writer(os.path.join(os.getcwd(), defaults.general.outdir, "txt2vid-samples"), im, extension=".mp4", fps=30)
|
||||
try:
|
||||
video_path = os.path.join(os.getcwd(), defaults.general.outdir, "txt2vid-samples","temp.mp4")
|
||||
writer = imageio.get_writer(video_path, fps=24)
|
||||
for frame in frames:
|
||||
writer.append_data(frame)
|
||||
writer.close()
|
||||
except:
|
||||
print("Can't save video, skipping.")
|
||||
|
||||
# show video preview on the UI
|
||||
st.session_state["preview_video"].video(open(video_path, 'rb').read())
|
||||
|
||||
mem_max_used, mem_total = mem_mon.read_and_stop()
|
||||
time_diff = time.time()- start
|
||||
|
||||
info = f"""
|
||||
{prompts}
|
||||
Sampling Steps: {num_steps}, Sampler: {scheduler}, CFG scale: {cfg_scale}, Seed: {seeds}, Max Frames: {max_frames}""".strip()
|
||||
stats = f'''
|
||||
Took { round(time_diff, 2) }s total ({ round(time_diff/(max_frames),2) }s per image)
|
||||
Peak memory usage: { -(mem_max_used // -1_048_576) } MiB / { -(mem_total // -1_048_576) } MiB / { round(mem_max_used/mem_total*100, 3) }%'''
|
||||
|
||||
return im, seeds, info, stats
|
||||
|
||||
#
|
||||
def layout():
|
||||
with st.form("txt2vid-inputs"):
|
||||
st.session_state["generation_mode"] = "txt2vid"
|
||||
|
||||
input_col1, generate_col1 = st.columns([10,1])
|
||||
with input_col1:
|
||||
#prompt = st.text_area("Input Text","")
|
||||
prompt = st.text_input("Input Text","", placeholder="A corgi wearing a top hat as an oil painting.")
|
||||
|
||||
# Every form must have a submit button, the extra blank spaces is a temp way to align it with the input field. Needs to be done in CSS or some other way.
|
||||
generate_col1.write("")
|
||||
generate_col1.write("")
|
||||
generate_button = generate_col1.form_submit_button("Generate")
|
||||
|
||||
# creating the page layout using columns
|
||||
col1, col2, col3 = st.columns([1,2,1], gap="large")
|
||||
|
||||
with col1:
|
||||
width = st.slider("Width:", min_value=64, max_value=2048, value=defaults.txt2vid.width, step=64)
|
||||
height = st.slider("Height:", min_value=64, max_value=2048, value=defaults.txt2vid.height, step=64)
|
||||
cfg_scale = st.slider("CFG (Classifier Free Guidance Scale):", min_value=1.0, max_value=30.0, value=defaults.txt2vid.cfg_scale, step=0.5, help="How strongly the image should follow the prompt.")
|
||||
seed = st.text_input("Seed:", value=defaults.txt2vid.seed, help=" The seed to use, if left blank a random seed will be generated.")
|
||||
#batch_count = st.slider("Batch count.", min_value=1, max_value=100, value=defaults.txt2vid.batch_count, step=1, help="How many iterations or batches of images to generate in total.")
|
||||
#batch_size = st.slider("Batch size", min_value=1, max_value=250, value=defaults.txt2vid.batch_size, step=1,
|
||||
#help="How many images are at once in a batch.\
|
||||
#It increases the VRAM usage a lot but if you have enough VRAM it can reduce the time it takes to finish generation as more images are generated at once.\
|
||||
#Default: 1")
|
||||
|
||||
st.session_state["max_frames"] = int(st.text_input("Max Frames:", value=defaults.txt2vid.max_frames, help="Specify the max number of frames you want to generate."))
|
||||
|
||||
with st.expander("Preview Settings"):
|
||||
st.session_state["update_preview"] = st.checkbox("Update Image Preview", value=defaults.txt2vid.update_preview,
|
||||
help="If enabled the image preview will be updated during the generation instead of at the end. \
|
||||
You can use the Update Preview \Frequency option bellow to customize how frequent it's updated. \
|
||||
By default this is enabled and the frequency is set to 1 step.")
|
||||
|
||||
st.session_state["update_preview_frequency"] = st.text_input("Update Image Preview Frequency", value=defaults.txt2vid.update_preview_frequency,
|
||||
help="Frequency in steps at which the the preview image is updated. By default the frequency \
|
||||
is set to 1 step.")
|
||||
with col2:
|
||||
preview_tab, gallery_tab = st.tabs(["Preview", "Gallery"])
|
||||
|
||||
with preview_tab:
|
||||
#st.write("Image")
|
||||
#Image for testing
|
||||
#image = Image.open(requests.get("https://icon-library.com/images/image-placeholder-icon/image-placeholder-icon-13.jpg", stream=True).raw).convert('RGB')
|
||||
#new_image = image.resize((175, 240))
|
||||
#preview_image = st.image(image)
|
||||
|
||||
# create an empty container for the image, progress bar, etc so we can update it later and use session_state to hold them globally.
|
||||
st.session_state["preview_image"] = st.empty()
|
||||
|
||||
st.session_state["loading"] = st.empty()
|
||||
|
||||
st.session_state["progress_bar_text"] = st.empty()
|
||||
st.session_state["progress_bar"] = st.empty()
|
||||
|
||||
#generate_video = st.empty()
|
||||
st.session_state["preview_video"] = st.empty()
|
||||
|
||||
message = st.empty()
|
||||
|
||||
with gallery_tab:
|
||||
st.write('Here should be the image gallery, if I could make a grid in streamlit.')
|
||||
|
||||
with col3:
|
||||
# If we have custom models available on the "models/custom"
|
||||
#folder then we show a menu to select which model we want to use, otherwise we use the main model for SD
|
||||
#if CustomModel_available:
|
||||
custom_model = st.selectbox("Custom Model:", defaults.txt2vid.custom_models_list,
|
||||
index=defaults.txt2vid.custom_models_list.index(defaults.txt2vid.default_model),
|
||||
help="Select the model you want to use. This option is only available if you have custom models \
|
||||
on your 'models/custom' folder. The model name that will be shown here is the same as the name\
|
||||
the file for the model has on said folder, it is recommended to give the .ckpt file a name that \
|
||||
will make it easier for you to distinguish it from other models. Default: Stable Diffusion v1.4")
|
||||
|
||||
#st.session_state["weights_path"] = custom_model
|
||||
#else:
|
||||
#custom_model = "CompVis/stable-diffusion-v1-4"
|
||||
#st.session_state["weights_path"] = f"CompVis/{slugify(custom_model.lower())}"
|
||||
|
||||
st.session_state.sampling_steps = st.slider("Sampling Steps", value=defaults.txt2vid.sampling_steps, min_value=10, step=10, max_value=500,
|
||||
help="Number of steps between each pair of sampled points")
|
||||
st.session_state.num_inference_steps = st.slider("Inference Steps:", value=defaults.txt2vid.num_inference_steps, min_value=10,step=10, max_value=500,
|
||||
help="Higher values (e.g. 100, 200 etc) can create better images.")
|
||||
|
||||
#sampler_name_list = ["k_lms", "k_euler", "k_euler_a", "k_dpm_2", "k_dpm_2_a", "k_heun", "PLMS", "DDIM"]
|
||||
#sampler_name = st.selectbox("Sampling method", sampler_name_list,
|
||||
#index=sampler_name_list.index(defaults.txt2vid.default_sampler), help="Sampling method to use. Default: k_euler")
|
||||
scheduler_name_list = ["klms", "ddim"]
|
||||
scheduler_name = st.selectbox("Scheduler:", scheduler_name_list,
|
||||
index=scheduler_name_list.index(defaults.txt2vid.scheduler_name), help="Scheduler to use. Default: klms")
|
||||
|
||||
beta_scheduler_type_list = ["scaled_linear", "linear"]
|
||||
beta_scheduler_type = st.selectbox("Beta Schedule Type:", beta_scheduler_type_list,
|
||||
index=beta_scheduler_type_list.index(defaults.txt2vid.beta_scheduler_type), help="Schedule Type to use. Default: linear")
|
||||
|
||||
|
||||
#basic_tab, advanced_tab = st.tabs(["Basic", "Advanced"])
|
||||
|
||||
#with basic_tab:
|
||||
#summit_on_enter = st.radio("Submit on enter?", ("Yes", "No"), horizontal=True,
|
||||
#help="Press the Enter key to summit, when 'No' is selected you can use the Enter key to write multiple lines.")
|
||||
|
||||
with st.expander("Advanced"):
|
||||
st.session_state["separate_prompts"] = st.checkbox("Create Prompt Matrix.", value=defaults.txt2vid.separate_prompts,
|
||||
help="Separate multiple prompts using the `|` character, and get all combinations of them.")
|
||||
st.session_state["normalize_prompt_weights"] = st.checkbox("Normalize Prompt Weights.",
|
||||
value=defaults.txt2vid.normalize_prompt_weights, help="Ensure the sum of all weights add up to 1.0")
|
||||
st.session_state["save_individual_images"] = st.checkbox("Save individual images.",
|
||||
value=defaults.txt2vid.save_individual_images, help="Save each image generated before any filter or enhancement is applied.")
|
||||
st.session_state["save_video"] = st.checkbox("Save video",value=defaults.txt2vid.save_video, help="Save a video with all the images generated as frames at the end of the generation.")
|
||||
st.session_state["group_by_prompt"] = st.checkbox("Group results by prompt", value=defaults.txt2vid.group_by_prompt,
|
||||
help="Saves all the images with the same prompt into the same folder. When using a prompt matrix each prompt combination will have its own folder.")
|
||||
st.session_state["write_info_files"] = st.checkbox("Write Info file", value=defaults.txt2vid.write_info_files,
|
||||
help="Save a file next to the image with informartion about the generation.")
|
||||
st.session_state["dynamic_preview_frequency"] = st.checkbox("Dynamic Preview Frequency", value=defaults.txt2vid.dynamic_preview_frequency,
|
||||
help="This option tries to find the best value at which we can update \
|
||||
the preview image during generation while minimizing the impact it has in performance. Default: True")
|
||||
st.session_state["do_loop"] = st.checkbox("Do Loop", value=defaults.txt2vid.do_loop,
|
||||
help="Do loop")
|
||||
st.session_state["save_as_jpg"] = st.checkbox("Save samples as jpg", value=defaults.txt2vid.save_as_jpg, help="Saves the images as jpg instead of png.")
|
||||
|
||||
if GFPGAN_available:
|
||||
st.session_state["use_GFPGAN"] = st.checkbox("Use GFPGAN", value=defaults.txt2vid.use_GFPGAN, help="Uses the GFPGAN model to improve faces after the generation. This greatly improve the quality and consistency of faces but uses extra VRAM. Disable if you need the extra VRAM.")
|
||||
else:
|
||||
st.session_state["use_GFPGAN"] = False
|
||||
|
||||
if RealESRGAN_available:
|
||||
st.session_state["use_RealESRGAN"] = st.checkbox("Use RealESRGAN", value=defaults.txt2vid.use_RealESRGAN,
|
||||
help="Uses the RealESRGAN model to upscale the images after the generation. This greatly improve the quality and lets you have high resolution images but uses extra VRAM. Disable if you need the extra VRAM.")
|
||||
st.session_state["RealESRGAN_model"] = st.selectbox("RealESRGAN model", ["RealESRGAN_x4plus", "RealESRGAN_x4plus_anime_6B"], index=0)
|
||||
else:
|
||||
st.session_state["use_RealESRGAN"] = False
|
||||
st.session_state["RealESRGAN_model"] = "RealESRGAN_x4plus"
|
||||
|
||||
st.session_state["variant_amount"] = st.slider("Variant Amount:", value=defaults.txt2vid.variant_amount, min_value=0.0, max_value=1.0, step=0.01)
|
||||
st.session_state["variant_seed"] = st.text_input("Variant Seed:", value=defaults.txt2vid.seed, help="The seed to use when generating a variant, if left blank a random seed will be generated.")
|
||||
st.session_state["beta_start"] = st.slider("Beta Start:", value=defaults.txt2vid.beta_start, min_value=0.0001, max_value=0.03, step=0.0001, format="%.4f")
|
||||
st.session_state["beta_end"] = st.slider("Beta End:", value=defaults.txt2vid.beta_end, min_value=0.0001, max_value=0.03, step=0.0001, format="%.4f")
|
||||
|
||||
if generate_button:
|
||||
#print("Loading models")
|
||||
# load the models when we hit the generate button for the first time, it wont be loaded after that so dont worry.
|
||||
#load_models(False, False, False, RealESRGAN_model, CustomModel_available=CustomModel_available, custom_model=custom_model)
|
||||
|
||||
# run video generation
|
||||
image, seed, info, stats = txt2vid(prompts=prompt, gpu=defaults.general.gpu,
|
||||
num_steps=st.session_state.sampling_steps, max_frames=int(st.session_state.max_frames),
|
||||
num_inference_steps=st.session_state.num_inference_steps,
|
||||
cfg_scale=cfg_scale,do_loop=st.session_state["do_loop"],
|
||||
seeds=seed, quality=100, eta=0.0, width=width,
|
||||
height=height, weights_path=custom_model, scheduler=scheduler_name,
|
||||
disable_tqdm=False, beta_start=st.session_state["beta_start"], beta_end=st.session_state["beta_end"],
|
||||
beta_schedule=beta_scheduler_type)
|
||||
|
||||
#message.success('Done!', icon="✅")
|
||||
message.success('Render Complete: ' + info + '; Stats: ' + stats, icon="✅")
|
||||
|
||||
#except (StopException, KeyError):
|
||||
#print(f"Received Streamlit StopException")
|
||||
|
||||
# this will render all the images at the end of the generation but its better if its moved to a second tab inside col2 and shown as a gallery.
|
||||
# use the current col2 first tab to show the preview_img and update it as its generated.
|
||||
#preview_image.image(output_images)
|
||||
|
||||
|
||||
#on import run init
|
||||
def createHTMLGallery(images,info):
|
||||
html3 = """
|
||||
<div class="gallery-history" style="
|
||||
display: flex;
|
||||
flex-wrap: wrap;
|
||||
align-items: flex-start;">
|
||||
"""
|
||||
mkdwn_array = []
|
||||
for i in images:
|
||||
try:
|
||||
seed = info[images.index(i)]
|
||||
except:
|
||||
seed = ' '
|
||||
image_io = BytesIO()
|
||||
i.save(image_io, 'PNG')
|
||||
width, height = i.size
|
||||
#get random number for the id
|
||||
image_id = "%s" % (str(images.index(i)))
|
||||
(data, mimetype) = STImage._normalize_to_bytes(image_io.getvalue(), width, 'auto')
|
||||
this_file = in_memory_file_manager.add(data, mimetype, image_id)
|
||||
img_str = this_file.url
|
||||
#img_str = 'data:image/png;base64,' + b64encode(image_io.getvalue()).decode('ascii')
|
||||
#get image size
|
||||
|
||||
#make sure the image is not bigger then 150px but keep the aspect ratio
|
||||
if width > 150:
|
||||
height = int(height * (150/width))
|
||||
width = 150
|
||||
if height > 150:
|
||||
width = int(width * (150/height))
|
||||
height = 150
|
||||
|
||||
#mkdwn = f"""<img src="{img_str}" alt="Image" with="200" height="200" />"""
|
||||
mkdwn = f'''<div class="gallery" style="margin: 3px;" >
|
||||
<a href="{img_str}">
|
||||
<img src="{img_str}" alt="Image" width="{width}" height="{height}">
|
||||
</a>
|
||||
<div class="desc" style="text-align: center; opacity: 40%;">{seed}</div>
|
||||
</div>
|
||||
'''
|
||||
mkdwn_array.append(mkdwn)
|
||||
|
||||
html3 += "".join(mkdwn_array)
|
||||
html3 += '</div>'
|
||||
return html3
|
||||
File diff suppressed because it is too large
Load Diff
2738
scripts/webui_streamlit_old.py
Normal file
2738
scripts/webui_streamlit_old.py
Normal file
File diff suppressed because it is too large
Load Diff
Loading…
Reference in New Issue
Block a user