Color Match

Color Match transfers the color palette of a reference image onto a target image - the classic color-grading task that the post-processing layer keeps rediscovering. It's a port of KJNodes' ColorMatchV2 (reordered so the target image is the first input, which matters when you bypass the node: it passes the correct image through). If you've ever generated an image and thought "the colors are off," this is the honest alternative to re-rolling the generation: a statistics transfer is deterministic and instant, a re-gen is neither.

The three jobs this was made for: making an inpainted or outpainted region agree with the original, making an upscaled tile match its neighbors, and seating a composited subject into a new background's light.

How it works

Eclipse's version is unusually method-rich, and the method you pick is the whole ballgame:

• CPU methods (via the color-matcher library): mkl, hm, reinhard, mvgd, hm-mvgd-hm, hm-mkl-hm. These are the classic mean/variance and histogram-transfer algorithms, run per-frame on the CPU. Inputs are normalized to finite RGB values first for numerical stability, and if a result comes back invalid the node falls back to the target image rather than poisoning your batch.
• GPU reinhard_lab_gpu (Kornia): the Reinhard mean/std transfer done in LAB space on the GPU.
• GPU wavelet (Haar wavelet, LAB): decomposes both images, matches the low-frequency color band's statistics, and keeps the target's detail bands - so it preserves texture while transplanting color. This is the one to pick when you want the palette without the blur.
• GPU scattersort (exact histogram matching): sorts each channel and maps rank positions, producing an exact per-channel color distribution match. The most "total" of the bunch.

strength blends the result (0 = no change, 1 = full transfer, up to 10). multithread parallelizes CPU batch work; per_frame processes one frame at a time to cap VRAM on big batches.

The inputs that matter

• image - the target (the one being recolored). Passes through on bypass.
• image_ref - the reference whose colors get transferred.
• method - defaults to mkl; switch to wavelet for detail-preserving, scattersort for exact matching.
• strength - the blend amount.
• Output - image, the matched result.

Watch the hidden dependencies

Here's where people get burned. The CPU methods need the color-matcher package, and the Kornia GPU methods need kornia - and neither is in the pack's requirements.txt. If you pick mkl (the default!) on a fresh install, you may hit an ImportError that tells you exactly what to do:

pip install color-matcher   # for mkl / hm / reinhard / mvgd / hm-* CPU methods
pip install kornia          # for reinhard_lab_gpu and wavelet

scattersort is pure torch and needs neither. Install the missing package into the same Python that runs ComfyUI (for portable installs, python_embeded\python.exe -m pip install …).

Installing the node

Part of ComfyUI_Eclipse, installed once:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse

Restart ComfyUI (or ComfyUI Manager → search "ComfyUI Eclipse").

Gotchas

Pack-wide: ComfyUI_Eclipse was formerly RvTools, and v4.0.0 removed all legacy nodes - old workflows need the Workflow Migration Tool node (or python tools/migrate_workflow.py <workflow.json>) to rewrite old IDs with a backup. And remember: if a frame's color transfer produces invalid values, the node silently returns the original target for that frame - so a partially-unmatched result usually means the reference was degenerate (flat color, all-black), not that the node failed.