TS Color Match

Ever shot two videos on different cameras, or generated a batch where every frame has its own white balance? TS Color Match fixes that without touching a sampler. You feed it a reference image whose colors you like and a target batch that doesn't match it, and it transfers the palette over. Same idea as Photoshop's "Match Color," except it runs inside the graph, handles whole video batches, and doesn't lose its mind over the borders.

This is the color-grading node you actually reach for when the job is consistency, not look: harmonizing clips from different sources, matching CG into live-action plates, or grading a video to one keyframe you already approved. It's deterministic pixel math, not a model, so it's fast and it never rewrites your image - the post-processing layer's whole trick is doing in milliseconds what a diffusion pass would gamble a seed on.

How it works

Two algorithms, and the choice is a real trade, not marketing. MKL (the default) solves a fast linear color transfer and adds a temporal_smoothing flag that EMA-smooths the transform between frames - that's the one you want for video, because it kills flicker. Sinkhorn is optimal-transport based: slower, hungrier, but noticeably more precise when you're matching a dramatic grade. Start with MKL, and only reach for Sinkhorn when the MKL result looks off and you can afford the wait.

It works by computing a color transform from the reference image's statistics, then applying it to the target batch. To keep 4K sane it computes on a downscaled copy - compute_max_side (default 1024) is where you save VRAM, and dropping it to 512 costs almost nothing visually. There's also chunk_size for processing long videos in chunks rather than all at once, and a logging toggle if you want to watch it work.

The inputs that matter

• reference / target - the two IMAGE inputs. Reference is one image (or a batch where each frame gets its own stats); target is everything that gets recolored.
• mode - mkl or sinkhorn. Video → MKL. Precision → Sinkhorn.
• match_mask + mask_size - set to rectangle or ellipse and it samples only the border strip of width mask_size. Counterintuitive, but this is the stabilization trick for video: sampling the edges keeps the transform from swinging when the subject moves through the frame.
• reuse_reference - with a single reference and a whole batch, it computes the stats once and reuses them. This is the video-speed button.
• strength - blends the corrected result back into the original. 1.0 is full transfer; 0.5 gives you half the grade. Great for when the reference is close but not that close.

The single output is image - the recolored batch - which wires straight into a preview, saver, or the next node in the chain.

Installing it

It's part of comfyui-timesaver (search "Timesaver" in ComfyUI Manager), or:

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

Then restart ComfyUI. No model files - this node is pure math, so it works the moment the pack loads.

Where people get burned

The OOM trap is mkl_sample_points. At 0 it uses every pixel, which at 4K is a lot of matrix; the tooltip recommends 200k–500k for a reason. Same story with sinkhorn_max_points - keep it at 1024–2048 for 4K input. And if you're on a Mac, don't panic when you see CPU activity: the README notes this node quietly does two small linear-algebra solves on the CPU on Metal machines, because PyTorch doesn't implement linalg.eigh/linalg.lstsq on the GPU there. Everything else stays on the GPU, and the result is identical.