Replies: 1 comment 1 reply
-
Any ideas? Maybe I'll try reinstalling everything from scratch. |
Beta Was this translation helpful? Give feedback.
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I'm using AbdBarho/stable-diffusion-webui-docker with the Automatic1111 profile and it's always worked great. Recently I pulled and updated the container image and since then ControlNet inpainting no longer works like it's supposed to for me.
I noticed that the "upload mask" has been replaced with "effective region mask" and I've read the github issues about it, but I can't see why that would make a difference. My masks are created the same way they've always been: take an image into gimp, select the portion I want to replace, fill with white, then invert the selection and fill with black, resulting a black background with only the parts I want to inpaint completely white. The mask and the image are the same size. I select controlnet unit 0, enable, select Inpaint as the control type, pixel perfect, and effective region mask, then upload the image into the left and the mask into the right preview. Upon generation though, it's like there's no mask at all: I end up with an image identical to the original input image. If I used the cursor to "paint" the mask on the image it does seem to work, but this is a much less precise way of doing it. I've tried it with both inpaint_only and inpaint_only+lama preprocessors, with control_v11p_sd15_inpaint_fp16 and control_lora_rank128_v11p_sd15_inpaint_fp16 models, but nothing seems to make a difference.
Anybody have any ideas? I can do img2img inpainting, but I've always found controlnet inpaint to be better.
here's a screenshot of what used to work but is no longer working:
and here's a screenshot of it working using the "draw mask" brush, but that's not what I want to do. I want to upload the mask.
Beta Was this translation helpful? Give feedback.
All reactions