[Bug]: "Restore faces" does not work on different stable diffusion model

[E-16]

Is there an existing issue for this?

• [X] I have searched the existing issues and checked the recent builds/commits

What happened?

When I use a different stable diffusion model, "restore faces" feature does not work. I generating image with: "Save a copy of image before doing face restoration" It's very easy to see the difference.

When I using the model: sd-v1-4.ckpt [7460a6fa] everything works fine, the difference is easily visible.

But when I switch on a different model, for example, I use: anime-diffusion.ckpt [925997e9] two absolutely identical pictures come out (they are same even by the number of bytes). I tested with and without .vae.pt modification, no difference.

The guess is maybe face detector cant detect the face.

I can say that this was a problem even on a very old version, without hesitation can say this one for sure: b8c38f2bbfa28904f67f0c4f9cabab4d85ebced2

Steps to reproduce the problem

1. Switch Stable Diffusion checkpoint to anime-diffusion.ckpt [925997e9] (in my case)
2. Check "Restore faces" (for a quick check, with "Save a copy of image before doing face restoration" setting on)

What should have happened?

The output should be slightly different pictures, one of which has a restored face.

Commit where the problem happens

f49c08ea566385db339c6628f65c3a121033f67c

What platforms do you use to access UI ?

Windows

What browsers do you use to access the UI ?

Google Chrome

Command Line Arguments

Additional information, context and logs

Tried with argument: --use-cpu gfpgan But got an error: Scrolling through more than 40 similar bug reports, I found very similar bug signs, but mine is different in the last arguments: torch.cuda.FloatTensor and torch.FloatTensor. Tried mess with diffenent settings (changing versions, using different versions of venv, code modification) to solve this problem, but couldn't do anything. But my goal wasnt use cpu at all, I just check it.

[E-16]

So, I done some a bit research, test this issue on a different machine, on a recent commit 1ef32c8b8fa3e16a1e7b287eb19d4fc943d1f2a5 and the problem stay the same. So far I figure that .vae.pt modification as well as different or none hypernetworks does not affect the original model: sd-v1-4.ckpt [7460a6fa], with different configurations, "Restore faces" works fine.

Seems like the problem puerly in different stable diffusion model, I need to check more stable diffusion models then, to see how they will work or not.

[Modesty226]

Restore faces isn't working for me on v1-4.ckpt [7460a6fa] either. It's not working in text2img, img2img or Extras. All my faces are completely scrambled right now.

[ckombo]

I'm having the same problem. When I include Restore Faces, I get no output and this error: 'TypeError: 'NoneType' object is not subscriptable

[nophuxyz]

I'm having the same problem. When I include Restore Faces, I get no output and this error: 'TypeError: 'NoneType' object is not subscriptable

Exactly this for me too

[panchitooo000]

I'm having the same problem. When I include Restore Faces, I get no output and this error: 'TypeError: 'NoneType' object is not subscriptable

Did you figure out how to fix this?

[jhogsett]

I had and seemed to fix a similar issue just now (not getting a type error). Not sure if this is helpful:

I just installed from scratch on a new PC (Alienware R15) and new GPU (RTX 3090). I copied my "models" folder from my older PC (AW R11 w/ RTX 2080 TI) over to the new install to try to save time downloading. I noticed all faces were garbled despite choosing Codeformer or GFPGAN in the settings. I then tried toggling those settings, restarting, different resolutions and CFG scales and nothing worked.

I went to the Wiki troubleshooting page and followed the recommendation to "reinstall from scratch" (deleting "venv" and "repositories" folders. I also deleted all the models I had copied over except the stable diffusion ones (1.5, 2.0). I also disabled xformers (removed the cmdline option - can't be sure this made any difference).

On restarting, I am now getting perfect faces like on my old PC.

Edit: I re-enabled --xformers and it's still working OK.