Color Match 🆎

Every outpaint or inpaint eventually hits the same wall: the patched region comes back a hair brighter, a touch cooler, or just off compared to the plate around it. Re-rolling the seed is the wrong tool - that's the thing the knowledge base keeps hammering, and it's right: a statistics transfer is deterministic and instant, a regeneration is neither. Color Match 🆎 is the deterministic fix.

It transfers the reference image's color statistics onto your input. The default lab method moves per-channel mean and standard deviation in LAB space, which is the perceptual-friendly, "make this look like that" default. rgb does the same math in raw channels, mkl maps the full covariance (better when hues are rotated rather than just shifted), and histogram matches each channel's distribution exactly - the strongest, least subtle option. For the classic use case, lab at strength 1.0 is where you start.

The two inputs that make it useful

reference is the image whose colors you want to copy - for an inpainted crop, feed it the original untouched image. It broadcasts across a batch, so one reference serves a whole video's worth of frames.

mask is where this node gets clever, and it's also the thing that trips people up. The mask restricts both the statistics measurement and the correction to its white area - perfect for recoloring just an inpainted region instead of the whole frame. Black pixels pass through bit-identical. Here's the catch: the node has no mask output. That's not a bug. The mask only scopes the fix; it doesn't change your mask, so you keep wiring your original mask onward and don't need it echoed back. invert_mask flips which side gets corrected, and strength blends between the original and the fully matched result - 0.3 is a gentle nudge, 1.0 is a full takeover.

The video trick

For video there's reference_mode: first_frame, which locks every frame to the batch's own first frame as the target. That's the one-node flicker fix: instead of each frame being matched against a static reference, they're all pulled toward a single anchor so colors stop breathing across the clip. No reference input needed.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ausboss/ComfyUI-AusBoss.git

Restart ComfyUI and search for AusBoss (also in Manager under the pack title). No extra Python dependencies; the pack uses Pillow and NumPy that ship with ComfyUI. Minimum ComfyUI is 0.27.1.

Troubleshooting

If the match is too aggressive, that's what strength is for - a feathered inpaint seam usually wants 0.5–0.8, not 1.0. If hues look rotated rather than shifted, switch lab to mkl. And if your mask correction is landing on the wrong side of the frame, you've got invert_mask backwards, which is the most common mistake here. One real limitation: a plain statistics transfer can't fix a fundamentally different white balance or lighting angle - that's a regrade, not a match.