ComfyUI-layerdiffuse

ComfyUI implementation of https://github.com/layerdiffusion/LayerDiffuse.

Installation

Download the repository and unpack into the custom_nodes folder in the ComfyUI installation directory.

Or clone via GIT, starting from ComfyUI installation directory:

cd custom_nodes
git clone git@github.com:huchenlei/ComfyUI-layerdiffuse.git

Run pip install -r requirements.txt to install python dependencies. You might experience version conflict on diffusers if you have other extensions that depends on other versions of diffusers. In this case, it is recommended to setup separate Python venvs.

Workflows

Generate foreground

Generate foreground (RGB + alpha)

If you want more control of getting RGB image and alpha channel mask separately, you can use this workflow.

Blending (FG/BG)

Blending given FG

Blending given BG

Extract FG from Blended + BG

Extract BG from Blended + FG

Forge impl's sanity check sets Stop at to 0.5 to get better quality BG. This workflow might be inferior comparing to other object removal workflows.

Extract BG from Blended + FG (Stop at 0.5)

In SD Forge impl, there is a stop at param that determines when layer diffuse should stop in the denosing process. In the background, what this param does is unapply the LoRA and c_concat cond after a certain step threshold. This is hard/risky to implement directly in ComfyUI as it requires manually load a model that has every changes except the layer diffusion change applied. A workaround in ComfyUI is to have another img2img pass on the layer diffuse result to simulate the effect of stop at param.

Generate FG from BG combined

Combines previous workflows to generate blended and FG given BG. We found that there are some color variations in the extracted FG. Need to confirm with layer diffusion authors on whether this is expected.

[2024-3-9] Generate FG + Blended given BG

Need batch size = 2N. Currently only for SD15.

[2024-3-9] Generate BG + Blended given FG

Need batch size = 2N. Currently only for SD15.

[2024-3-9] Generate BG + FG + Blended together

Need batch size = 3N. Currently only for SD15.

Note

Currently only SDXL/SD15 are supported. See https://github.com/layerdiffuse/sd-forge-layerdiffuse#model-notes for more details.
To decode RGBA result, the generation dimension must be multiple of 64. Otherwise, you will get decode error: