2.4.0

[mashb1t]

What's Changed

• Change settings tab elements to be more compact
• Add clip skip slider
• Add select for custom VAE
• Add new style "Random Style"
• Update default anime model to animaPencilXL_v310
• Add button to reconnect the UI after Fooocus crashed without having to configure everything again (no page reload required)
• Add performance "hyper-sd" (based on Hyper-SDXL 4 step LoRA)
• Add AlignYourSteps scheduler by Nvidia, see
• Add TCD sampler and scheduler (based on sgm_uniform)
• Add NSFW image censoring (disables intermediate image preview while generating). Set config value default_black_out_nsfw to True to always enable.
• Add argument --enable-describe-uov-image to automatically describe uploaded images for upscaling
• Add inline lora prompt references with subfolder support, example prompt: colorful bird <lora:toucan:1.2>
• Add size and aspect ratio recommendation on image describe
• Add inpaint brush color picker, helpful when image and mask brush have the same color
• Add automated Docker image build using Github Actions on each release, see available images
• Add full raw prompts to history logs
• Change code ownership from @lllyasviel to @mashb1t for automated issue / MR notification

Details

• feat: support download huggingface files from a mirror site by @chenxinlong in feat:support download huggingface files from a mirror site #2637
• chore: update interposer from v3.1 to v4.0 by @mashb1t in feat: update interposer from v3.1 to v4.0 #2717
• feat: add button to reconnect UI without having to reload the page by @mashb1t in feat: add button to reconnect UI without having to reload the page #2727
• feat: add optional model VAE select by @mashb1t in Add VAE select #2867
• feat: choose random style by @mashb1t in feat: choose random style #2855
• feat: update anime from animaPencilXL_v100 to animaPencilXL_v310 by @mashb1t in feat: update anime from animaPencilXL_v100 to animaPencilXL_v310 #2454
• refactor: rename label for reconnect button by @mashb1t in refactor: rename label for reconnect button #2893
• feat: add full raw prompt to history log by @docppp in Full prompt log #1920
• fix: use correct border radius css property by @khanvilkarvishvesh in fix: use correct border radius css property #2845
• fix: do not close meta tag in HTML header by @e52fa787 in fix: do not close meta tag in HTML header #2740
• feat: automatically describe image on uov image upload by @mashb1t in feat: automatically describe image on uov image upload #1938
• add nsfw image censoring via config and checkbox by @mashb1t in add nsfw image censoring via config and checkbox #958
• feat: add align your steps scheduler by @mashb1t in feat: add align your steps scheduler #2905
• feat: add support for lora inline prompt references by @cantor-set in Adding support to inline prompt references #2323
• feat: add tcd sampler and discrete distilled tcd scheduler based on sgm_uniform (same as lcm) by @mashb1t in feat: add tcd sampler #2907
• feat: add performance hyper-sd based on 4step LoRA by @mashb1t in feat: add performance hyper-sd based on 4step LoRA #2812
• fix: remove leftover code for hyper-sd testing by @mashb1t in fix: remove leftover code for hyper-sd testing #2959
• feat: optimize model management of nsfw image censoring by @mashb1t in feat: optimize model management of nsfw image censoring #2960
• feat: progress bar improvements by @mashb1t in feat: progress bar improvements #2962
• feat: inline lora optimisations by @mashb1t in feat: inline lora optimisations #2967
• feat: change code owner from @lllyasviel to @mashb1t by @mashb1t in feat: change code owner from @lllyasviel to @mashb1t #2948
• feat: only use valid inline loras, add subfolder support by @mashb1t in feat: only use valid inline loras, add subfolder support #2968
• feature: Reads the size and ratio of the image and gives the recommended size by @xhoxye in feature: Reads the size and ratio of the image and gives the recommended size #2971
• feat: build and push container image for ghcr.io, update docker.md, and other related fixes by @xynydev in feat: build and push container image for ghcr.io, update docker.md, and other related fixes  #2805
• feat: adjust line ending default config by @mashb1t in feat: adjust line ending default config #2991
• feat: add translation for image size describe by @mashb1t and @xhoxye in feat: add translation for image size describe #2992
• feat: read value 'CFG Mimicking from TSNR' from presets by @Alexdnk in feat: read value 'CFG Mimicking from TSNR' from presets #2990
• feat: add inpaint brush color picker by @mashb1t in feat: add inpaint brush color picker #2997
• feat: remove labels from most of the image input fields by @mashb1t in feat: remove labels from most of the image input fields #2998
• feat: add clip skip handling by @mashb1t in feat: add clip skip handling #2999
• feat: make ui settings more compact by @Alexdnk and @mashb1t in Slightly more compact ui settings #2590

Full Changelog: 2.3.1...v2.4.0

[ILikeToasters]

Video of almost all changes (someone snuck in the interface change last second...lol). Hopefully this will help explain some of the details for people if I didn't get anything wrong: https://youtu.be/8uqcdCF2Esk

[MeAmCat]

idk what you guys did, but my image quality somehow doubled. Nice work. I am confused about the Hyper-SD thingy tho. Is it enabled by default or where do I enable it?

Edit: nvm I'm stupid found it.

[DavidDragonsage]

@ILikeToasters So you "are" Kleebz Tech AI. Your channel is absolutely the best video source for Fooocus information! Thank you so much! ✨👍

[ILikeToasters]

That would be me. You are welcome and thanks for watching.

[BassAzayda]

@ILikeToasters So you "are" Kleebz Tech AI. Your channel is absolutely the best video source for Fooocus information! Thank you so much! ✨👍

Without a doubt!

[AnalyticsXXI]

That would be me. You are welcome and thanks for watching.

Oh, then I am your fun too xD. Just started to watch your guides about Fooocus, very good videos, you really understand how to explain things.

[cassiebug23]

For some reason my photos are coming out quite different with this new update. My custom checkpoints aren't as high quality as they were with the previous version/no longer fully resemble the people they were trained on. I was making pictures before it updated and things looked fine, so I'm not sure what's happened. All my settings are the same as they were pre-update.

[Krashlog]

Try adjusting the clip skip as I mentioned above. That might solve issue for moment.

unfortunately I did try that, it did not solve the issue for me.

It did get results much closer to the previous version, but the faces are still noticeably worse.

[ILikeToasters]

I figured it was CLIP skip since I did not encounter the last week. Also had an image I did generate real quick this morning when I did try the CLIP skip and It did recreate the one that I had set it to 2 but did not recreate the one I did not touch the slider and it just generated as if CLIP skip was on 2 since it was the same result(same seed). hope that makes sense.

[cassiebug23]

Try adjusting the clip skip as I mentioned above. That might solve issue for moment.

unfortunately I did try that, it did not solve the issue for me.

It did get results much closer to the previous version, but the faces are still noticeably worse.

I agree, it helped very slightly but the faces in my custom checkpoints and loras no longer look as close to the person as they normally do.

[Krashlog]

Sorry for the inconvenience and thank you for your valuable feedback, will keep you posted.

No problems, thanks for the hard work.👍The revert did take care of the problems for most of my gens, so that's a good place to start your investigation.

[mashb1t]

fixed in #3011, took a long time to figure out that Fooocus by default uses clip skip 2 and the functionality was correct, but the config & default values weren't. This is also the reason why clip skip 2 yielded the same results as before and why every image without clip skip was different.

[mzisa]

Hello, thanks for the update. I am curretnly facing this issue

Traceback (most recent call last):
File "D:\Fooocus_win64_2-1-831\Fooocus\entry_with_update.py", line 46, in
from launch import *
File "D:\Fooocus_win64_2-1-831\Fooocus\launch.py", line 136, in
config.default_base_model_name, config.checkpoint_downloads = download_models(
File "D:\Fooocus_win64_2-1-831\Fooocus\launch.py", line 127, in download_models
load_file_from_url(url=url, model_dir=config.paths_checkpoints[0], file_name=file_name)
File "D:\Fooocus_win64_2-1-831\Fooocus\modules\model_loader.py", line 27, in load_file_from_url
download_url_to_file(url, cached_file, progress=progress)
File "D:\Fooocus_win64_2-1-831\python_embeded\lib\site-packages\torch\hub.py", line 620, in download_url_to_file
u = urlopen(req)
File "urllib\request.py", line 216, in urlopen
File "urllib\request.py", line 525, in open
File "urllib\request.py", line 634, in http_response
File "urllib\request.py", line 563, in error
File "urllib\request.py", line 496, in _call_chain
File "urllib\request.py", line 643, in http_error_default
urllib.error.HTTPError: HTTP Error 404: Not Found

[mashb1t]

@mzisa make sure disable any proxy/VPN before testing and ensure the model URL exists if you have changed the code.
If this doesn't help please open a new issue and fill out all required fields + provide the URL of the model which can't be found.

[mzisa]

Thanks for your help. Actually I found it was due to my presets: it was trying to download the model but the download path was set to a file that belongs to my local filesystem. So no worries, not a bug, my bad

[DavidDragonsage]

@mashb1t - I really appreciate all the hard work you put into this, and it also great to see Fooocus back on track again! I also appreciate the work of everyone else involved in this release. Thank you all! 🎂🥳👍

[badraymen]

hi, i'm not expert en coding, but how can i come back to the 2.3 version, i have the same issues :(

[mashb1t]

@badraymen make sure to do one of the following:

• restart Fooocus to get the latest version including the clip skip fix
• use the version without fix (2.4.0) and set clip skip to 2 for identical handling as before
• manually switch to 2.3.1

I could not notice any speed decrease, rather a bump instead due to clip skip and/or hyper-sd.
Make sure you don't accidentally have 2 Fooocus instances open and close all irrelevant applications running in the background.

[badraymen]

@mashb1t
I am extremely sorry, yes it was my fault I apologize, I launched fooocus and SDWUI at the same time to use "Clip interrogator" the Prompt generator.
Please a suggestion: is there a possibility of adding "Clip interrogator" to "Fooocus", it would be really great, especially since the extension goes wonderfully with SDXL
Thank you for the remarkable effort and the speed of problem solving and really all my respect to all the people who participated in creating this marvel

[mashb1t]

@badraymen Fooocus already has an image describe functionality using BLIP, which you can find in the image input > describe tab.
There are no plans of including other image interrogation / description tools but you're welcome to open a feature request.

[badraymen]

@mashb1t with all respect for your efforts to guarantee the user the best practical way to better use your magnificent software, I lived the experience with the "describe" tool but the prompts are very limited, on the other hand I would be happy with the fact of added this extension to fooocus if I know how to launch a suggestion or a ticket

[mashb1t]

@badraymen => https://github.com/lllyasviel/Fooocus/issues/new/choose

[MantisOW]

yesterday I had problems with my faceswaps with inpaint.... and I didn't even have the usual results... is there any way I can change the settings back or the version?

[mashb1t]

@MantisOW the issue has been resolved, please restart Fooocus to get the automated update.

[MantisOW]

thank you, will try this and hope for the best 👍

[bobgus39]

what happens with gradio's live server? it stops working right now.

Thanks.

[AnalyticsXXI]

what happens with gradio's live server? it stops working right now.

Yes, you can check it here https://status.gradio.app/
Usually such things lasted couple of hours.