Auto Color Drift Correction V4.3 (两遍校正)

If you've ever generated a Wan or SCAIL-2 video in segments - and for anything longer than 81 frames, you have to - you know the tell. The first five frames of every new chunk are a slightly different color temperature than the tail of the previous one. It's not noise, it's drift: each sampling pass lands on marginally different color values, and when you stitch the segments together the seams flash. Auto Color Drift Correction exists to kill exactly that.

It's the V4.3 "two-pass" version of the author's color stabilizer, and it's one of the more thoughtful nodes in this pack because it stops pretending drift is a simple offset. The name is descriptive: it detects a drift, corrects the whole new segment globally, then goes back and micro-adjusts the residual per frame. No API, no model download, nothing to fetch - it's pure tensor math on the frames you feed it.

How it works

Feed it the current segment's frames as images and it treats the first overlap_count frames (default 5) as the overlap with the previous segment, leaving them untouched. Everything after that gets corrected in passes:

• Pass 0 (prev_frames alignment): if you wire in the previous segment's tail via prev_frames, the whole segment is shifted to match it. This is the precise seam alignment you want in a real multi-segment loop.
• Pass 1 (global offset): it averages the last few overlap frames against the first few new frames, computes the jump vector, and applies one clamped offset to all new frames. Per-channel clamping to max_offset (default 0.02, roughly 5.1 pixel values) keeps it from overcorrecting.
• Pass 2 (residual fine-tune): each new frame is pulled toward the reference independently, clamped to max_offset × residual_strength. Because each frame is computed on its own, errors don't accumulate down the segment the way an EMA chain would.

In auto mode the node decides for you: if the multi-frame jump is below a threshold (0.0005), it decides there's no real drift and just passes through. That's the mode to start in - it's genuinely the "set it and forget it" option.

The inputs you'll actually set

• mode - standard (you tune), auto (it decides), off (pure bypass). Start with auto.
• overlap_count - must match how many frames of your previous segment you actually kept as overlap.
• prev_frames - the previous segment's tail. Wiring this is the difference between "good" and "seamless".
• seam_strength, max_offset, residual_strength - the dials. If you see banding, lower residual_strength; if the seam is still obvious, raise seam_strength.

Outputs are corrected_images (the fixed batch, ready to feed your save/encode node) and drift_info, a string that logs every offset it applied in readable form plus a JSON block. Keep an eye on it once - if the numbers stay tiny, the node is mostly along for the ride, which is exactly what you want.

Installing it

This ships in user2318/ComfyUI-CustomNodeKit, a Chinese-authored toolkit built around WanAnimate and SCAIL-2 long-video workflows. Easiest install is ComfyUI Manager: search "CustomNodeKit" and hit install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/user2318/ComfyUI-CustomNodeKit.git
cd ComfyUI-CustomNodeKit
pip install -r requirements.txt

then restart ComfyUI. The requirements.txt pulls in tqdm, opencv-python, mediapipe and imageio-ffmpeg, and the pack's install.py auto-installs groundingdino-py and transformers on startup for the object-detection nodes. This node itself is pure torch - zero extra weight.

The gotcha

It only helps if you actually put it in the loop. The drift you're fixing happens between segments, so this needs to sit inside your segment loop (after VAE decode, before the next segment's context window reads the tail) with prev_frames connected. Sticking it on a single already-generated video does nothing useful, and the README's own workflow files treat it as part of the segment pipeline rather than a post-process. Wire it like that and the flashing seams just stop.