沐阳 H3 · 段间漂移校正

Long chains of generated segments have a smell problem. Each link in the continuity chain is conditioned on the previous one, so small biases - a bit too dark, a touch too saturated, a slight color cast - accumulate until segment 14 looks like it was shot through a different lens than segment 2. The README is honest that a clean seam doesn't guarantee stable content down the chain, and 沐阳 H3 · 段间漂移校正 (H3DriftCorrect) is the optional corrective for exactly that: it aligns this segment's color statistics to an anchor segment.

The smart part is which anchor. The default align mode is 衔接处 (seam): it only compares a handful of frames either side of the cut. Here's the reasoning, and it's genuinely good: the last frames of the previous segment and the first frames of this one show almost the same moment in time - so whatever differs between them is the model's accumulated bias, not a real scene change. Compare whole clips instead and you'd read an intentional change of scenery as drift and "correct" it right out. The whole-segment mode exists for when you want the entire segment's distribution pulled toward the anchor (segment 1), but the tooltip warns it hits hard whenever the scene changes.

It's also gentle by design: the node fits one transform for the whole segment and applies it uniformly, so you get a stable correction rather than per-frame flicker. method defaults to mean_std (per-channel) with an MKL option that also handles rotation in color space; strength at 1.0 fully aligns, 0.5 goes halfway if you want to keep some natural inter-segment variation.

Inputs

• images - this segment, after sampling.
• anchor - the alignment target: the previous segment in seam mode, segment 1 in whole mode.
• method - mean_std (default) or MKL.
• align - 衔接处 (seam, recommended) or whole-segment.
• strength - 0.0–1.0, default 1.0.
• sample_frames - how many frames each side of the cut to compare in seam mode (default 8).

Outputs: images (corrected) and report (STRING) - a text summary showing saturation/brightness before and after versus the anchor. That report is the debugging gift: it tells you whether the correction actually moved the numbers.

Install and where it fits now

Pack install: ComfyUI Manager search "ComfyUI-MiniMaxH3-Myang", or git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang into custom_nodes, restart. No extra Python deps.

One honest note about its place in the pack: the long-video node used to carry drift correction built-in, and the changelog shows it was pulled out - the current design prefers latent continuity over color correction, so this node is now a standalone you add to a manually-wired graph when the numbers drift. Wire it between sampling and the seam/trim step, read the report, and let the numbers tell you whether it's helping. And the pack-wide reminder: H3's local weights are territory-restricted by community license (US/EU/UK/South Korea excluded).