Using the model from: diffusers/stable-diffusion-xl-1.0-inpainting-0.1 · Hugging Face
PRE-INSTALLED PYTHON REQUIRED
I recommend you install 3.10 or higher, I have 3.10.6 just FYI. Older versions may work, but I can't confirm that.
This is a small Gradio GUI that allows you to use the diffusers SDXL Inpainting Model locally. It comes with some optimizations that bring the VRAM usage down to 7-9GB, depending on how large of an image you are working with. I recommend using the "EulerDiscreteScheduler". This GUI is similar to the Huggingface demo, but you won't have to wait for your turn and there are also additional features. Use the browser zoom (CTRL + Mouse Wheel) to inpaint with better accuracy.
v2: Now with integrated outpainting, mask inversion, noise mask, and mask blur
Installation:
Just run "sdxl_inpainting_installer.bat", the cmd window should close automatically once it is finished, after which you can run "sdxl_inpainting_launch.bat" (the first time will take quite a while because it is downloading the inpainting model from Huggingface) or the "no_ops" version if you have the VRAM but it will use ~10GB for just a regular 1024x1024 image (pretty pointless if you ask me because I did not observe a significant slow-down with VRAM optimizations). The GUI does not automatically open, you have to copy-paste the local IP into your browser.
Description
FAQ
Comments (14)
The config attributes {'decay': 0.9999, 'inv_gamma': 1.0, 'min_decay': 0.0, 'optimization_step': 37000, 'power': 0.6666666666666666, 'update_after_step': 0, 'use_ema_warmup': False} were passed to UNet2DConditionModel, but are not expected and will be ignored. Please verify your config.json configuration file.
I don't see any config files to change. I'm not sure if that's an error but when I inpaint something it barely changes the image. This is what it does for me on any sdxl inpainting so it may be my end.
I get this error as well, but it does not cause an issue for me. From what I know those are just training parameters and they should not matter for generation.
Regarding the image issue, what parameters are you using? If you set the inference steps or the denoising strength too low, it is only going to change a tiny bit.
I set it just like yours, 30-35 steps, 99 denoise and various other settings and still doesn't change much, like ill put large orange pumpkin or something and it will just change the hand angle or something. It could be my prompts or something because everyone else doesn't seem to have a problem with sdxl inpainting.
@milksteak are you experiencing the same problems on here?: SDXL Inpainting - a Hugging Face Space by diffusers
@massOxygen Yep, it just resized the image on HF
@milksteak huh that is weird, sounds like an issue with the browser for me, because the HF demo uses its own GPU.
@massOxygen Awesome thanks, it's working in firefox but I had tried chrome and brave incognito and thought that wasn't it. Amazing little GUI, thanks again!
This creates his own Env, it will not mess up with other dependencies or Envs. right?
That was my intention, it creates its own virtual environment so I hope it does not mess with the "main" one.
@massOxygen Working fine, where are the images stored?
Also, it only works with 0.99 Denoise value, if i try a less than that i get error.
@massOxygen How we update it?
I don't see any update.bat
@zerocool22 I released a new version.
@massOxygen missing a noise image: std_noise_image_1,png it gives Errno 2 and closes the window
@xaid thank you for pointing that out! It should be fixed now.