Awesome
ComfyUI-IC-Light-Native
ComfyUI native implementation of IC-Light.
Install
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-IC-Light-Native.git
Download models
IC-Light main repo is based on diffusers. In order to load it with UnetLoader in ComfyUI, state_dict keys need to convert to ldm format. You can download models with ldm keys here: https://huggingface.co/huchenlei/IC-Light-ldm/tree/main
There are 2 models:
- iclight_sd15_fc_unet_ldm: Use this in FG workflows
- iclight_sd15_fbc_unet_ldm: Use this in BG workflows
After you download these models, please put them under ComfyUI/models/unet
and load them with UNETLoader
node.
Recommended nodes
- ComfyUI-KJNodes: Provides various mask nodes to create light map.
- ComfyUI-Easy-Use: A giant node pack of everything. The remove bg node used in workflow comes from this pack.
- ComfyUI_essentials: Many useful tooling nodes. Image resize node used in the workflow comes from this pack.
- ComfyUI-IC-Light: The IC-Light impl from kijai. It includes a very useful
DetailTransfer
node to help preverse high frequency details from input fg image.
Workflows
Please make sure the fg image's masked/transparent area are grey before you pass it to the VAE. Otherwise, you will get background obscured in FC workflows or
darkened background in FBC workflows. You can use IC Light Apply Mask Grey
to make sure the masked area's color is correct. See following examples:
Given FG, Generate BG and relight
If you want to keep the original color of the fg object, you can put the fg object in the latent space to further guide the generation. workflow
Given FG and light map, Genereate BG and relight
Light from right
Light from left
Given FG and BG, Put FG on BG and relight
Recover high frequency detail (Text, etc) from original input image
- Input image:
- Raw generation:
- After detail transfer:
Common Issues
IC-Light's unet is accepting extra inputs on top of the common noise input. FG model accepts extra 1 input (4 channels). BG model accepts 2 extra input (8 channels). The original unet's input is 4 channels as well.
If you see following error, it means you are using FG workflow but loaded the BG model.
RuntimeError: Given groups=1, weight of size [320, 8, 3, 3], expected input[2, 12, 64, 64] to have 8 channels, but got 12 channels instead
If you see following error, it means you are using FG workflow but loaded the BG model.
RuntimeError: Given groups=1, weight of size [320, 12, 3, 3], expected input[2, 8, 64, 64] to have 12 channels, but got 8 channels instead