Added the Txt2Img settings from the `configs/webui/webui_streamlit.yaml` file. All are working except `separate_prompts`, which throws a `Missing key separate_prompts` error for some reason.
# Checklist:
- [x] I have changed the base branch to `dev`
- [x] I have performed a self-review of my own code
- [x] I have commented my code in hard-to-understand areas
- [x] I have made corresponding changes to the documentation
* Restore --config. This will be useful when you have an init config
that you don't want overwritten.
* Cache the individual transformed images in TextualInversionDataset.
This gains speed by avoiding reading and reprocessing the image each
time it's used for training.
* Turn on no_grad for inference and clean up tensors during
checkpointing. This reduces memory usage slightly.
* Set the sample output size to 384x384. We just need them large enough
for manual evaluation, and this gains us a decent chunk of speed.
* (breaking change) Custom templates are now semicolon-delineated.
Additionally, custom templates are properly passed through to
TextualInversionDataset to generate input_ids for your images. Using
custom templates which accurately describe your input images seems to
improve training fidelity.
* Cache autoencoding of image pixel data. This substantially increases
the speed of training, upwards of 40% for me.
* Clean up a little bit of cruft.
There was a safety check where RealESRGAN and loopback
were not allowed to be on at the same time to prevent
rapidly growing images (I think, at least) which was
causing confusion in the UI since there was no indication
this wasn't allowed. Using ESRGAN just on the final iteration
should be safe, so this commit enables that.
- Improved txt2vid so its now possible to generate high resolution images with less VRAM than before.
- Added condition to ensure that the pipe model from txt2vid is removed from memory when switching from the txt2vid tab to txt2img and viceversa.
* Some options on the Streamlit txt2img page now follow the defaults from the relevant config files.
* Fixed a copy-paste gone wrong in my previous commit.
* st.session_state["defaults"] fix
Co-authored-by: hlky <106811348+hlky@users.noreply.github.com>
This is a collection of several changes to enhance image display:
* When using GFPGAN or RealESRGAN, only the final output will be
displayed.
* In batch>1 mode, each final image will be collected into an image grid
for display
* The image is constrained to a reasonable size to ensure that batch
grids of RealESRGAN'd images don't end up spitting out a massive image
that the browser then has to handle.
* Additionally, the progress bar indicator is updated as each image is
post-processed.
The regex was not accounting properly for prompt weights that didn't begin with a leading number such as .5 or .1 and was instead splitting those off into their own prompt which got everything all screwed up.
For example, the prompt string of "Fruit:1 grapes:-.5" should parse as
[('Fruit', 1.0), ('grapes', -.5)]
but was being incorrectly parsed as
[('Fruit', 1.0), ('grapes', 1.0), ('-.5', 1.0)]
This fixes that by making the regex properly catch decimals.
* JobManager: Re-merge #611
PR #611 seems to have got lost in the shuffle after
the transition to 'dev'.
This commit re-merges the feature branch. This adds
support for viewing preview images as the image
generates, as well as cancelling in-progress images
and a couple fixes and clean-ups.
* JobManager: Clear jobs that fail to start
Sometimes if a job fails to start it will get stuck in the active job
list. This commit ensures that jobs that raise exceptions are cleared,
and also adds a start timer to clear out jobs that fail to start
within a reasonable amount of time.
- Fixed several issues with wrong variable being used on multiple places.
- Addd option to toggle some extra option from the config file for when the model is loading on the txt2vid tab.
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
- Added Dynamic Preview Frequency option for the txt2vid tab which tries to find the lowest value for update_preview_frequency at which we can update the preview image during generation while at the same time minimizing the impact it has in performance.
- Added option to save a video file on the outputs/txt2vid-samples folder after the generation is complete similar to how the save_grid option works on other tabs.
- Added a video preview which shows a video on the txt2vid tab when the generation is completed.
- Formated some lines of code to make it use less space and fit on the a single screen.
- Added a script called Settings.py to the script folder in which Settings for the Setting page will be placed. Empty for now.
- Improved txt2vid speed by 2 times.
- Added DDIM scheduler.
- Added sliders for beta_start and beta_end to have more control over these parameters on the scheduler.
- Added option to select the scheduler type from scaled_linear or linear.
- Added option to save info files for the txt2vid tab and improved the information saved to include most of the parameters used to run the generation.
- You can now download any model from the huggingface website to use on the txt2vid tab, just add the name to the custom_models_list on the config file.
* webui: display the GPU in use during startup
tell the user which GPU the code is actually going to use before spending lots of time loading everything onto the GPU
* typo
* add some info messages
* evaluate current GPU properly
* add debug flag gating
not everyone wants or needs to see debug messages :)
* add in stray debug msg
* webui: detect scoped-down GPU environment
check if we're using a scoped-down GPU environment (pynvml does not listen to CUDA_VISIBLE_DEVICES) so that we can measure memory on the correct GPU
* remove unnecessary import
* Perform masked image restoration when using GFPGAN or RealESRGAN, fixing #947.
Also fixes bug in image display when using masked image restoration with RealESRGAN.
When the image is upscaled using RealESRGAN the image restoration can not use the
original image because it has wrong resolution. In this case the image restoration
will restore the non-regenerated parts of the image with an RealESRGAN upscaled
version of the original input image.
Modifications from GFPGAN or color correction in (un)masked parts are also restored
to the original image by mask blending.
* Update scripts/webui.py
Co-authored-by: Thomas Mello <work.mello@gmail.com>
color correction is already used for loopback to prevent color drift with the first image as correction target.
the option allows to use the color correction even without loopback mode.
it helps keeping the colors similar to the input image.
* Add mask_restore option to give users the option to restore images based on mask, fixing #665.
Before commit c73fdd78 (Implement masking during sampling to improve blending, #308)
image mask was applied after sampling, resulting in masked parts that are not regenerated
to actually stay the same.
Since c73fdd78 the masked img2img will change the whole image, even in masked areas.
It gives better looking results at first glance, but will result in image degredation when
applied a few times. See issue #665.
In the workflow of using repeated masked img2img, users may want to use this options to keep the parts
of image they actually want to keep without image degradation. A final masked img2img or whole image img2img with mask_restore disabled
will give the better blending of "Implement masking during sampling".
* revert changes of a7be43ba in change_image_editor_mode
* fix ui_functions.change_image_editor_mode by adding gr.update to the end of the list it returns
* revert inserted newlines and whitespaces to match format of previous code
* improve caption of new option mask_restore
"Only modify regenerated parts of image"
* fix ui_functions.change_image_editor_mode by adding gr.update to the end of the list it returns
an old copy of the function exists in webui.py, this superflous function mistakenly was changed by the earlier commit b6a9e16b
* remove unused functions that are near duplicates of functions in ui_functions.py
* Metadata cleanup - Maintain metadata within UI
This commit, when combined with Gradio 3.2.1b1+, maintains image
metadata as an image is passed throughout the UI. For example,
if you generate an image, send it to Image Lab, upscale it, fix faces,
and then drag the resulting image back in to Image Lab, it will still
remember the image generation parameters.
When the image is saved, the metadata will be stripped from it if
save-metadata is not enabled. If the image is saved by *dragging*
out of the UI on to the filesystem it may maintain its metadata.
Note: I have ran into UI responsiveness issues with upgrading Gradio.
Seems there may be some Gradio queue management issues. *Without* the
gradio update this commit will maintain current functionality, but
will not keep meetadata when dragging an image between UI components.
* Move ImageMetadata into its own file
Cleans up webui, enables webui_streamlit et al to use it as well.
* Fix typo
