Skip to content

Latest commit

 

History

History
109 lines (68 loc) · 6.39 KB

README.md

File metadata and controls

109 lines (68 loc) · 6.39 KB

Replacer

Replacer is an extension for AUTOMATIC1111/stable-diffusion-webui. The goal of this extension is to automate objects masking by detection prompt, using sd-webui-segment-anything, and img2img inpainting in one easy to use tab. It also useful for batch inpaint, and inpaint in video with AnimateDiff

You also can draw your mask instead of or in addition to detection, and take advantage of convenient HiresFix option, and ControlNet inpainting with preserving original image resolution and aspect ratio

If you find this project useful, please star it on GitHub!

Installation

  1. Install sd-webui-segment-anything extension. If it bothers you, you can hide it in the Replacer's settings. Go to tab Extension -> Available -> click Load from and search "sd-webui-segment-anything"
  2. Download model sam_hq_vit_l.pth, or others from the list bellow, and put it into extensions/sd-webui-segment-anything/models/sam
  3. For faster hires fix, download lcm-lora-sdv1-5, rename it into lcm-lora-sdv1-5.safetensors, put into models/Lora. Or if you have already lcm lora, then change hires suffix in the extension options
  4. Install this extension. Go to tab Extension -> Available -> click Load from and search "Replacer". For AMD and Intel GPUs, and maybe something else, you need to enable CPU for detection in Replacer's settings. Go to Settings -> Replacer and enable it
  5. Reload UI

If you don't want to use Video feature, that's all for you. Further steps are for Video:

  1. Install sd-webui-animatediff and sd-webui-controlnet extensions. You should also use Extension -> Available tab and find them there
  2. Download mm_sd15_v3.safetensors animatediff's motion model, and put it into extensions/sd-webui-animatediff/model directory
  3. Download control_v11p_sd15_inpaint_fp16.safetensors controlnet's model and put it into models/ControlNet directory
  4. I strongly recommend you to download mm_sd15_v3_sparsectrl_rgb.safetensors and mm_sd15_v3_sparsectrl_scribble.safetensors controlnet's models. Put them also into models/ControlNet directory. Then you can select SparseCtrl module in ControlNet extension. The rgb one requires "none" preprocessor
SAM models list:

SAM-HQ are the best for me. Choose it depending on your vram. Sum this model size with dino model size (694MB-938MB)

  1. SAM from Meta AI.

  2. SAM-HQ from SysCV.

  3. MobileSAM from Kyung Hee University.

FastSAM and Matting-Anything aren't currently supported

How does it work?

First, grounding dino models detect objects you provided in the detection prompt. Then segment anything model generates contours of them. And then extension chooses randomly 1 of 3 generated masks, and inpaints it with regular inpainting method in a1111 webui

When you press the "Apply hires fix" button, the extension regenerates the image with exactly the same settings, excluding upscaler_for_img2img. Then it applies inpainting with "Original" masked content mode and lower denoising but higher resolution.

Supported extensons:

  1. Lama cleaner as masked content

  2. Image Comparison

  3. ControlNet

  4. AnimateDiff

  5. ar-plusplus (maybe works with "Aspect Ratio and Resolution Buttons" forks)

  6. Other extension scripts which doesn't have control arguments, e.g. Hardware Info in metadata, NudeNet NSFW Censor, built-in Hypertile

Docs:


Need to do:

  • ☑️ cache mask
  • ☑️ batch processing
  • ☑️ "apply hires fix" button
  • ☑️ additional options
  • ☑️ progress bar + interrupt
  • ☑️ option for pass into hires fix automatically
  • ☑️ control net
  • ☑️ pass previous frame into ControlNet for video
  • tiled vae
  • ☑️ "hide segment anything extension" option
  • ☑️ txt2img script
  • more video and mask input types
  • RIFE frames interpolation
  • allow multiply instances (presets)

small todo:

  • add hires fix args into metadata
  • FreeInit support