MiniMax H3 Flow-Aligned Regenerate

Here's the workflow pain this whole pack exists to fix: generate H3 at a low resolution, upscale with a learned latent upscaler, then refine at full resolution - and watch the refine pass cheerfully re-walk the entire denoising path from scratch, half of which you already paid for and already got right. MiniMax H3 Flow-Aligned Regenerate is the node that makes that second pass remember the first one. It patches your high-resolution model so the refine run steers its estimates toward what the low-res run already decided was good.

It's the explicit two-pass version of the pack's guidance, for people who keep the classic low-gen → learned-upscale → refine structure rather than migrating to the Continuum-integrated refine path (that's Flow-Aligned Refine State's job). Mechanism, honestly explained: at each denoising point on the high-res run, the node looks up the matched low-res state in the trajectory you captured in pass one, and nudges the high-res predicted-clean estimate toward it - only in low spatial frequencies, controlled by low_frequency_cutoff, and decaying as the high-res stage progresses. It's a correction in the direction of the trustworthy low-res answer, not a copy of it.

Inputs that matter. model is the high-res H3 model, trajectory is the shared handle (the one you already captured with), and guidance_mode is where you'll spend your time. direction is the conservative default and the current recommendation - everything else in the list (direction+acceleration, direction+temporal, downsample_consistency) is implemented, instrumented, and functionally validated but has not shown a decoded-media win in the author's matched tests. The weights that go with the mode: direction_weight (0.35 default here, stronger than the progressive nodes' 0.25), acceleration_weight, consistency_weight, temporal_weight - leave the non-default ones at zero unless you're deliberately experimenting. The author is refreshingly blunt that these extra modes are "research," not "better."

Two optional conditioning inputs deserve a mention because they're the safety mechanism: source_conditioning and source_negative. The trajectory identity includes a bounded conditioning fingerprint, so guidance only fires when the second pass provably descends from the first. Connect them to keep that check honest (and you'll get an error if you feed source_negative without source_conditioning). Outputs: patched model into your high-res sampler, plus metrics for the shared benchmark artifact.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/MiniMax-H3-Flow-Aligned-Regenerate.git

Restart. No Python deps beyond what ComfyUI ships, no model downloads - but you do need the trajectory infrastructure running: a Flow Trajectory handle plus Trajectory Capture on pass one. This is a guidance adapter over H3, not a standalone generator.

Where people get burned

The whole trick collapses if pass one and pass two don't share one trajectory handle in the same graph run - you'll get fails-closed errors rather than silent corruption, which is the good outcome. And this node does not call any MiniMax API or reproduce the closed H3-Regenerate-2K model; it's an independent training-free approximation, and the author says so in the README. Decoded media, not telemetry, is the quality gate - so when you're testing modes, actually look at the clips. Guidance RMS ratios on a chart are not the same thing as motion that doesn't smear.