Load And Apply IC-Light
Turn any SD1.5 checkpoint into a relighting model
- model
- MODEL
Every IC-Light graph starts here. This is the node that takes a normal SD1.5 checkpoint you already have loaded and grafts lllyasviel's relighting weights onto it, so the same model that was drawing you portraits five seconds ago is now consuming a foreground image and putting a new light source on it. Skip this node and nothing downstream in the pack does anything - it's the load-bearing first step, not an optional extra.
IC-Light itself is a bit of a legend at this point: released in May 2024 by lllyasviel (also behind ControlNet and Forge), the code hasn't changed since the day after launch, and it's still the go-to for one specific job - relighting where the subject can't be allowed to drift. It works by training the model to treat light transport as linear and consistent, which forces it to change illumination while leaving material and structure alone. That's why a good IC-Light pass looks like a render instead of a colour-overlay filter. kijai wrote this ComfyUI wrapper in the release-week thread, essentially on a dare ("shouldn't be too hard to add natively into Comfy too, looking into it") - it became the standard way almost everyone runs the model.
How it works
IC-Light doesn't ship as a full standalone checkpoint you load fresh - it's weights meant to be applied on top of an SD1.5 UNet you supply. This node does the applying: hand it a model from your usual checkpoint loader, tell it which IC-Light file to use, and it comes back as a patched MODEL ready for relighting. Handy detail from the pack's own docs: it works with any SD1.5 checkpoint, not just base SD1.5 - so if you've got a favorite realistic or stylized 1.5 model, you don't have to give it up to use IC-Light.
The inputs and outputs that matter
model(MODEL) - your regular SD1.5 checkpoint, loaded the normal way, wired straight in.model_path- a dropdown of IC-Light.safetensorsfiles it finds on disk. There are three worth knowing:iclight_sd15_fcis the default (foreground + text prompt),iclight_sd15_fconis the same thing trained with offset noise (reach for it iffclooks washed out in the shadows), andiclight_sd15_fbcadditionally takes a background image as lighting context - simpler to prompt, but it tends to bleed background colour into the subject, sofcoften composites better even when you have a background to work with.
Output is a single MODEL - feed it to your KSampler and pair it with ICLightConditioning to actually supply the foreground/lighting.
How to install it
Through ComfyUI Manager is the recommended path - search "IC-light" and it pulls in the node pack with no extra requirements. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-IC-Light, restart.
The node pack is only half the job, though - the weights are a separate download. Manager can grab those too (search "IC-light" again, under models this time) and by default they land in ComfyUI/models/unet/IC-Light. Or pull them straight from huggingface.co/lllyasviel/ic-light and drop them anywhere under ComfyUI/models/unet. All three checkpoints are Apache-2.0, so no license friction.
Common issues & troubleshooting
The model_path dropdown is empty. This trips up almost everyone on first use - the node reads whatever's in models/unet, and if you only installed the custom node pack (not the weights), there's nothing to list. Go download at least fc from the HuggingFace repo above.
It errors or produces garbage on a non-SD1.5 model. IC-Light's weights are shaped for SD1.5's UNet specifically. The project genuinely froze in May 2024 - there's no SDXL or Flux version, and there won't be; a Flux-based V2 was announced but the weights were never released, so this is what exists. Feed it an SD1.5 checkpoint, not your SDXL or Flux model.
Results look flat or washed out. Try fcon instead of fc - same model, trained with offset noise, and it's the one to reach for when fc's shadows feel muddy.
You picked fbc expecting cleaner backgrounds and got weird colour casts instead. That's the known tradeoff - fbc pulls colour off the background into your subject. Unless you specifically need the background-image-as-light-source workflow, fc plus a text description of the lighting usually blends better.
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| model_path | COMBO | 0 options: |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |