Fooocus Advanced 2

[lllyasviel]

This is a continued thread of "Fooocus Advanced".
All discussions should continue here.
Previous: Fooocus Advanced 1

[lllyasviel]

Hi everyone,

A bit late, but we want to let you know that 2.1.771 just fixed the bugs of GPT generating unwanted objects.

2.1.771 BUG Fix: GPT should not generate unwanted objects.

Prompt: "stone", seed 12345, 16 images, 2.1.761 (with bug, before fix)

many unwanted objects, a few completely off-topic.

Prompt: "stone", seed 12345, 16 images, 2.1.771 (after fix)

16/16 images are stone.

Prompt: "apple", seed 12345, 16 images, 2.1.761 (with bug, before fix)

many unwanted objects, a few completely off-topic.

Prompt: "apple", seed 12345, 16 images, 2.1.771 (after fix)

16/16 images are apple.

Prompt: "knife", seed 12345, 16 images, 2.1.761 (with bug, before fix)

many unwanted objects, a few completely off-topic.

Prompt: "knife", seed 12345, 16 images, 2.1.771 (after fix)

16/16 images are knife.

Prompt: "wall", seed 12345, 16 images, 2.1.761 (with bug, before fix)

many unwanted objects, a few completely off-topic.

Prompt: "wall", seed 12345, 16 images, 2.1.771 (after fix)

16/16 images are wall.

BUG Influence

version	influence
1.x (including forks)	influenced
2.x ~ 2.1.761 (including forks)	influenced
2.1.761 ~ 2.1.769 (including forks)	highly influenced
2.1.771 ~ now (including forks)	fixed, not influenced

Side Note

Note that this BUG is not introduced by recently updates. It is a long-standing problem and recently it becomes more severe. Fortunately, the problem is permanently solved in after 2.1.771. If you want other objects, please directly write down them in prompt and they will be generated. The long-sentence prompt understanding is also improved in many of our recent tests.

Besides, all our future improvements will be in line with state-of-the-art models and all best results from CivitAI. If you are stilling using the previous sd_xl_base_1.0_0.9vae.safetensors default config (!!deprecated and not recommended!!), we highly enourage users to delete user_path_config.txt to refresh Fooocus, or directly re-download Fooocus and let Fooocus to configure/download newer default models.

Finally, if you are fork maker or distribution maker, do note that (1) the deprecated base+refiner are 13GB and will cause lots of waste of user internet traffic, download time, and difficulty in first launch; (2) newer config is only 6.5GB and quality are better than previous config in most cases; (3) many users have RAM problems, or System Swap problems, or HDD problems: they simply cannot launch Fooocus with previous config, or they waste several minutes everytime the refiner swap, when they accidentally put Fooocus on an old HDD and meet some RAM/swap problems.

[deadpipe]

Hi @lllyasviel

I liked the previous version which was more creative for small prompts. It helped to create varied images in batch generation.
Can you add an option where we can toggle between the new patch and the old version? It would be really helpful for people like me

[lllyasviel]

@deadpipe

we will have --chaos in the future for people who want random objects.
But the chaos will be set to 0 as default, same as MJ.
also you can reproduce previous results with previous logs. The format is

a dog
a dog, extremely XXX, XXX

where "a dog" is your prompt and "a dog, extremely XXX, XXX" is previous expansion in your log. You need to turn off current FooocusV2 to do this.

However, we do not recommend this: if you need some object, the better way is to directly type it. For example, instead of hoping that the random generator will give you a magic dog by coincident, you should just type

dark magic dog with magic effetcs

[putuoka]

@lllyasviel can u fix the hand & fingers please? also inpainting didn't work

[lllyasviel]

2.1.777
begin to support Ctrl+Up/Down to increase/decrease weights in prompts.
Both positive prompt and negative prompt can use this quick shortcut.

[bubuhu]

您好，我特别喜欢您的这个软件，实在是太棒了，但是我遇到一个问题，样式预览图一直加载不出来，很是头疼，重启很多次都不管用，有什么办法吗？

[lllyasviel]

2.1.782

2.1.782 is mainly an update for a new LoRA system that supports both SDXL loras and SD1.5 loras.

Now when you load a lora, the following things will happen:

1. try to load the lora to the base model, if failed (model mismatch), then try to load the lora to refiner (refiner can be SD1.5).
2. try to load the lora to refiner, if failed (model mismatch) then do nothing.

Note that in Fooocus, refiner can swap at 0.4 or 0.5 so you are basically using SD1.5 if your refiner is SD1.5.

In this way, Fooocus 2.1.782 can benefit from all models and loras from CivitAI with both SDXL and SD1.5 ecosystem, using the unique Fooocus swap algorithm, to achieve extremely high quality results (although the default setting is already very high quality), especially in some anime use cases, if users really want to play with all these things.

Recently the community also developed LCM loras. Users can use it by setting the sampler as 'lcm', sceduler as 'lcm', the forced overwrite of sampling step as 4 to 8, and CFG guidance as 1.0, in dev tools. Do not forget to change the LCM lora weight to 1.0 (many people forget this and report failure cases). Also, set refiner to None. If LCM's feedback in the artists community is good (not the feedback in the programmer community of Stable Diffusion), Fooocus may add some other shortcuts in the future.

[mrsunshine74]

If it helps the LCM community it works very well, I experimented for only an hour but I was able to cut the generation time in half while keeping the result pretty much the same, which I already find a very good result....CFG at 2 and even steps of 16, with a few steps 4-8 I didn't get a good result! While I have not yet been able to get lora sd 1.5 to work, either it gives me error or simply fooocus generates without using it...I will test tonight with more time available, if in any case there are others "at work" any advice is welcome! Thanks to all

[astro1df]

how do i change the number of steps in fooocus

[lllyasviel]

forced overwrite of sampling step below sampler @astro1df

[mrsunshine74]

Thanks for your work and this latest news is nothing short of crazy.... as for LCM it really looks like another bombshell, however we will experiment.Thank you very much!!!!

[lllyasviel]

2.1.785

• The user_path_config.txt is deprecated since 2.1.785. If you are using it right now, please use the new config.txt instead. See also the new documentation in the Readme.
• The paths in user_path_config.txt will still be loaded in recent versions, but it will be removed soon.
• We use very user-friendly method to automatically transfer your path settings from user_path_config.txt to config.txt and usually you do not need to do anything.
• The new config.txt will never save default values so the default value changes in scripts will not be prevented by old config files.

[lllyasviel]

2.1.788

• Inpaint engine v2.6 join the beta test of Fooocus inpaint models. Use it in dev tools -> inpaint engine -> v2.6 .

[mrsunshine74]

Guys I always have an error related to lora sd 1.5, has anyone been able to use it? It always gives me error (model mismatch)....I have tried various ways but nothing..... sorry in advance for my unfamiliarity .......

[lllyasviel]

make sure that file is not corrpted. otherwise post the link of that lora

[AkoZ]

hey i thought we can't use old SD1.5 loras ! .. also crashed too .. so to be tested now !

[lllyasviel]

2.1.790

• Face swap (in line with Midjourney InsightFace): Input Image -> Image Prompt -> Advanced -> FaceSwap
• The performance is super high. Use it carefully and never use it in any illegal things!
• This implementation will crop faces for you and you do NOT need to crop faces before feeding images into Fooocus. (If you previously manually crop faces from images for other software, you do not need to do that now in Fooocus.)

[xhoxye]

It is also related to the timing of the refiner model switch.

[django68]

I've already spent a long time experimenting with the Stop At and Weight controls as suggested. The issue isn't the last ten per cent losing the likeness, it seems to happen very early on in the steps. It looks better around 23 than the completed image, which is more of a very bad look alike. I've seen very good examples, so no doubt the problem is with me rather than face swap itself, but I don't see people saying what base models they are using sometimes, refiners settings.
I'm a newbie at this so still learning. Can you please explain and maybe suggest some refiner model switch settings to a begiiner. Thanks so much.

[thiner]

I've already spent a long time experimenting with the Stop At and Weight controls as suggested. The issue isn't the last ten per cent losing the likeness, it seems to happen very early on in the steps. It looks better around 23 than the completed image, which is more of a very bad look alike. I've seen very good examples, so no doubt the problem is with me rather than face swap itself, but I don't see people saying what base models they are using sometimes, refiners settings.
I'm a newbie at this so still learning. Can you please explain and maybe suggest some refiner model switch settings to a begiiner. Thanks so much.

Are you using refiner model?

[django68]

Thanks for the replies. I've tried with and without refiner. I do get some reasonable results, but even with those same settings, it throws out completely unrecognisable faces too. I'd love to see people with good results showing their chosen model, Stop At, Weight, Refiner setting etc. Thanks.

[code0wl]

I am a bit late to the party. Regarding the following feature.

This implementation will crop faces for you and you do NOT need to crop faces before feeding images into Fooocus. (If you previously manually crop faces from images for other software, you do not need to do that now in Fooocus.)

I was wondering if cropping could be done through the API prior to feeding it to Fooocus. Or is that also done automatically? From my results it looks like an UI only feature.

[lllyasviel]

2.1.799

• Added 'Extreme Speed' performance mode (based on LCM). The previous complicated settings are not needed now.

[mrsunshine74]

I can only congratulate you for what you do with Fooocus, really always one step ahead.....thank you very much!

[lllyasviel]

Recently some member of us reported several discovered spamming from some brand-new GitHub accounts with 0 repo/issue/star/stared/fork/follow/followed record (and some registered in less than one month). These spamming are removed and handled by GitHub TOS, but we will still take a look at it.

[lllyasviel]

2.1.802

Default inpaint engine changed to v2.6. You can still use inpaint engine v1 in dev tools.
Fix some VRAM problems.

[jepjoo]

Requesting a toggle for enabling/disabling facexlib face detection when using FaceSwap.

I was using IP-adapter Face previously by switching the file into the place of normal IP-adapter. Now that FaceSwap is a feature in Fooocus, I'm noticing that some images produce worse results with the added facexlib face detection added into the pipeline. In one case I'm seeing facexlib detect 2 faces in an image where there is very clearly only one face present.

IP-adapter Face works great even without facexlib so it would be nice if we could have a tickbox to choose whether to use facexlib or not. The added face detection also increases total generation time.

[lllyasviel]

In one case I'm seeing facexlib detect 2 faces in an image where there is very clearly only one face present.
show the image.
also a1111 used the same model for all face detections for months without problem

IP-adapter Face works great even without facexlib
No. the model is trained with face affine transform, LMEDS + some mask-like augmentations; and clipvision needs square

The added face detection also increases total generation time.
0.71 seconds on my CPU (do not even need GPU)

I'm noticing that some images produce worse results
show the image

[jepjoo]

This one gets a "Detected 2 faces".

-image edited out-

On the technical points I don't have much to say as I don't understand that stuff. But I have been getting good results by just replacing the standard IPa .bin file with IP-a Face .bin file. Example of such "wrong" usage from before FaceSwap was implemented, using the above image as input:

-image edited out-

[lllyasviel]

interesting image. I can reproduce the "detected 2 faces".
2.1.805 supports Skip Preprocessors in dev tools
However even in this "2 faces case", the crop still works and the correct face is cropped

[lllyasviel]

2.1.803

responsive ui

notepad_Q6I4t1rFjX.mp4

[astro1df]

can you please add support for these kind of lcm models
https://huggingface.co/latent-consistency/lcm-sdxl

[Yatching]

lllyasviel added an Extreme Speed mode in 2.1.799 update, which is based on lcm models. And work even better than on A1111 in my quick test.

[thiner]

I found that dogs always sit on floor, even I write the prompt "standing on xxx ", is this a bug?

[3Diva]

@thiner Yeah, that is strange. I prompted for "dog standing on coffee table" and it just ignored the dog part and just gave me coffee table:

To get a dog I had to prompt for "an animal standing on a coffee table, dog standing on a coffee table" and I had to add "sitting" to the negative prompt:

[thiner]

@3Diva good for you. I tried your prompt, it drew a standing dog, and plus a dog sitting on the floor. 🤣

[lllyasviel]

2.1.808

• Aspect ratios now show aspect ratios.
• Added style search.
• Added style sorting/ordering/favorites.

msedge_TtdRlRNZqR.-.Trim.mp4

[xhoxye]

a problems

[lllyasviel]

user hacked label of aspect ratio should be removed since they are not needed and does not support auto ratio computation @xhoxye

[fancellu]

Odd, works fine for me now. Maybe it was a bug. Very odd.

[lllyasviel]

Community Note:

Hi everyone,

Recently we see some fake websites in the name of Fooocus. They are not real and they are not from Fooocus.

Thanks!

[jony100200]

how do I use embeddings like this in focus?
Example

https://civitai.com/models/3036/charturner-character-turnaround-helper-for-15-and-21

[tralala87]

Any change to add option to upload multiple images at a time for Upscale instead of one at a time?

[deadpipe]

that would be very useful

[seoeaa]

Tell me, can I add https://huggingface.co/h94/IP-Adapter-FaceID

[thiner]

I didn't try, but I can see there are loras in the repo. Maybe you can just downlaod proper lora file (any of the .bin or .safetensors files), put it into models/loras folder in Fooocus, then mention the lora trigger words in prompt.
The FacdID author claimed a non-commercial license, I guess that's why it's not adopt in Fooocus offcially.

[adam2am]

Hey guys, is it possible to implement this in fooocus? https://github.com/TencentARC/PhotoMaker

[tre4b]

Quick quesiton. In the Image prompt each image area has an edit mode. It seems to allow you to choose an area of an image, is this intentional and does it do anything, or is this a future feature waiting to happen? I've looked but have not found mention of this anywhere.

[bkkjohn]

How can the refiner be enabled when running on a paid CoLab plan (with GPU support)?

[mashb1t]

@bkkjohn not quite, but almost: you can use any model as refiner, but most models are good enough + some better without selecting an additional refiner model.
Refiner and image prompt are features which use more resources than only image generation, which is why free tier instances on Colab (such as T4) struggle with these features without additional adjustments (see #1710 and don't ask me why it's been closed by the creator... ^^)
Hope the matter is now clear and easily teachable to others. All the best!

[bkkjohn]

Indeed, very useful! The SD instructions state: "The SDXL model is, in practice, two models. You run the base model, followed by the refiner model. The base model sets the global composition. The refiner model adds finer details. (You can optionally run the base model alone.)" So is it the refiner model that's part of SDXL disabled? Or does that option add a second refiner after?

And is there any way to turn the refiner option on when using a paid version of Colab?

[mashb1t]

Please don't get this wrong: Nothing is disabled on Colab, you can use StabilityAI's SDXL official with refiner by modifying the Colab command, adding --preset sai after entrypoint_with_update.py --share`.
The key fact is that by default the official SAI refiner isn't downloaded as none of the other presets use it.
This does not add a second refiner, it adds the one refiner. You will see that adding a refiner will enable you to select a refiner switch aka. the percentage after which the refimer is used instead of the base model when generating an image. The total amount of steps is not changed by using a refiner.

[bkkjohn]

That's a great resource! Is there no way to enable the refiner when running on CoLab, even if GPUs are available?

[mashb1t]

Nothing is disabled on Colab, you can use StabilityAI's SDXL official with refiner by modifying the Colab command, adding --preset sai after entrypoint_with_update.py --share`.

^ please read correctly. Nothing is disabled on Colab.
I am able to use a T4 instance (Colab Free) with refiner without problems.

!pip install pygit2==1.12.2
%cd /content
!git clone https://github.com/lllyasviel/Fooocus.git
%cd /content/Fooocus
!python entry_with_update.py --share --always-high-vram --preset sai

[madm1nds]

I want to express a huge thank you❤️‍🔥! Thank you for the work you have done! I never thought this could be run on 4gb vram😱.

I think your work is fantastic! Thanks❤️‍🔥.

[kesselerbr]

I'm still struggleing with face swap.
I've noticed a pattern, when you set a fixed seed, every 20 images, a similar person appears.
But still, faaaaar from the face I've settled to reproduce. I've tried my combo default, the reality fooocus default, jaugernauth...
Sometimes, seems it will get there. But failed.

I've tried faceswap+prompt/ faceswap+inpaint/faceswap+vary.
Am I missing something?

In dev tools, can I set to jump 20by20 instead of 1by1. Or even, force it to vary in the SAME SEED?
anyone is getting good results, or people just don't share the models ans LoRas they are working with?

Thank you lllyasviel
I'm having a really good time

[tornado73]

Good afternoon! thank you for your amazing work, tell me if it is possible to save used prompts in a png file, as is implemented in AUTOMATIC1111?
snx!

[mashb1t]

Metadata handling has been implemented in #1940, but it's not yet merged. Feel free to use my fork https://github.com/mashb1t/Fooocus if you want to have access to this feaure already. You can find instructions how to switch here.

[tornado73]

SNX!

[Jannchie]

I have some question about "Fooocus V2 Style".

This feature seems to just select some positive tags from a list. Is there any special task that definitely requires GPT-2 to complete?

Personally, I feel like a simple algorithm that randomly selects words could be used instead, without needing an additional GPT-2 model. This might be faster and save memory.

[mashb1t]

Fooocus v2 does not randomly select words from a list but applies basic logic to match words with content.

[jony100200]

what is the difference between foocus and the forge UI? foocus is a different image generator? Is there a guide on how to place all the models in one place to reduce the files that take place in my SDD?

[mashb1t]

Fooocus and Forge are two completely different tools, both using Stable Diffusion as AI basis.
Forge is built on top of A1111.

You can use a tool like https://github.com/LykosAI/StabilityMatrix or just use symlinks in the Fooocus config files to point to the existing models folder. Paths for model can be in array syntax. Hope this helps!

[jony100200]

Can Foocus create seamless textures to be used for 3D model surfaces like roofs etc? Is there any plugin that I can install in foocus to make seamless tiled textures? For example in A111 or forge UI I got the asymmetric tiling option(https://github.com/tjm35/asymmetric-tiling-sd-webui?tab=readme-ov-file)

[mashb1t]

@jony100200 this is still not possible, see last update at #2070

[jony100200]

is there any plans to implement this in future?

[mashb1t]

@jony100200 currently not, it's in the backlog with no planned release date.

[jony100200]

Hi, I like the generate prompt from the image feature in focus is it possible to batch process all the images in a given folder? Is it also possible to use those prompts to generate the pictures in order? I would also like to know if this feature can be turned into a simple app. I would like to see what prompts the image may produce. Also if there is already an existing feature or script like this please let me know thanks!

[mashb1t]

@jony100200 batch processing of a folder is currently not possible, but feel free to adjust the code as this only needs a few lines to be adjusted to your usecase.

Prompts can be read in order using a wildcard file and activating the checkbox "Read wildcards in order" (advanced > advanced > developer debug mode, see https://github.com/lllyasviel/Fooocus/blob/main/webui.py#L445).

To turn this feature into an app you need to manually call the code from your project and you need a device powerful enough to render the images. Please find the minimum system requirements in the readme or use cloud servers. If you want to integrate Fooocus to your app, you have to ensure to load and manage the whole process of generating the image or run it in the cloud + use the Websockets API.
No matter how you do it, keep in mind that Fooocus is licensed under GPLv3 and consult a lawyer if you ever want to use it commercially.

[Anistdev07]

Hi, there When i am runing the code this error prompt appears any reason behind this....?

[mashb1t]

Seems like you're not using Fooocus but DeFooocus. Please add a comment to ehristoforu#15 (same issue it seems).