Color Match Image

Run img2img at a high denoise, or relight something, or upscale through a second pass, and you'll notice the colors drift - a warmer cast here, a saturation bump there, nothing you asked for. Color Match Image exists to pull that back: give it a reference image with the color you actually want, and it nudges your image toward matching it.

It's the same family of fix that shows up constantly in relighting workflows - IC-Light in particular is notorious for shifting skin tones warm, and the community's standard fix for that is frequency separation: take color/tone from one source and detail from another, merged back together. This node gives you two ways to do that matching, from a crude global nudge to something closer to real frequency separation.

How it works

• images (IMAGE) - what you're correcting.
• reference (IMAGE) - the color you want it to match.
• blur_type - blur or guidedFilter. This decides how the match is computed.
• blur_size (default 0, 0–1023) - the key setting. At 0, the node matches on the overall mean color of both images - a blunt, whole-image shift. Push it up and both images get blurred by that radius first, so the match happens on regional/low-frequency color rather than a single global average - which is closer to true frequency separation and handles images where color varies across the frame (a face lit from one side, say) far better than a flat mean shift.
• factor (default 1, -10 to 10) - strength. 1 applies the full correction; lower values apply it partially; negative values push away from the reference instead of toward it, which is mostly useful for deliberately exaggerating a mismatch rather than fixing one.

Output is a single corrected IMAGE.

blur_type: guidedFilter is the edge-aware option - it respects boundaries in the image while computing the local color average, so you get less bleed across hard edges than plain gaussian blur at the same blur_size. It costs more compute; for most matching jobs plain blur is fine, and you reach for guidedFilter when you're seeing color bleed across a clear edge in the corrected result.

How to install it

ComfyUI Manager: search "ComfyUI-Image-Filters" → install → restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/spacepxl/ComfyUI-Image-Filters

then restart. The pack needs opencv-contrib-python; on Windows the author ships install.bat and an import_error_install.bat that wipes conflicting opencv variants and reinstalls the matched set - the same fix on Linux/manual setups is uninstalling stray opencv packages and installing opencv-contrib-python alone. Nothing to download otherwise - no models, negligible VRAM.

Common issues & troubleshooting

Global match (blur_size=0) looks wrong on an unevenly lit image. A mean-color shift can't fix regional color problems - if one side of your subject picked up a color cast and the other didn't, a single global mean will only get you halfway there or overcorrect the side that wasn't wrong. Raise blur_size so the match happens locally instead of as one flat number.

Result looks patchy or blotchy. Usually blur_size too small relative to the features you're matching - you're getting a match radius smaller than the color regions in the image, so neighboring patches get corrected differently and it shows. Push blur_size up until the patchiness smooths out.

Bleed across a hard edge (background color creeping onto the subject, or vice versa). Switch blur_type to guidedFilter - it's built exactly for keeping the local averaging from crossing strong edges.

Fixing IC-Light-style skin-tone shift specifically: this node alone gets you most of the way, but the community's more thorough fix layers this kind of low-frequency color match with a separate high-frequency detail pass from the original (this pack's Restore Detail or Median Filter Image nodes cover that half) - pure color matching won't recover lost fine detail, only tone.