MiniMax H3 Progressive Handoff

Two-pass H3 refinement wastes time because it repeats part of the trajectory. MiniMax H3 Progressive Handoff tries something sneakier: don't run two separate passes at all. Start sampling on the cheap low-resolution grid, then, partway through the same schedule, swap the video state up to the full target grid and keep going. One run, one handoff, no second low-sigma pass to pay for.

That's the pitch, and it's genuinely clever on paper. But you should know the framing up front: this node is the source-input variant - it assumes the surrounding workflow can legitimately start at source geometry and grow. If you're on Continuum, that assumption is false, and the README is explicit that you want the sibling "Progressive Handoff (Target Input)" instead, which keeps the workflow target-sized and hides the geometry change inside the wrapper. This node is for the cases where growing in place is fine.

How it works. The node wraps your H3 model so sampling runs on the smaller grid until handoff_coordinate (0.35 of the way through the 1→0 sigma schedule by default) - H3's flow coordinate, not a raw step count. At the boundary it does an exact low-grid probe, transfers the predicted-clean video state up to the target grid, rebuilds the target conditional state with deterministic noise, and continues with a forced actual evaluation. It also deliberately resets sampler and Spectrum feature history across the boundary, because cached transformer features from one spatial shape are meaningless at another. Audio is never spatially resized - H3's joint audio stays on its native path the whole time.

Inputs that matter. target_mode picks scale (a multiplier on the source, default 1.2) or pixels (explicit target_width/target_height). handoff_coordinate 0.35 with handoff_selection fixed is the tested operating point; auto_compute derives a geometry-aware handoff and works, but showed no quality advantage in the author's tests. Then the standard guidance menu (guidance_mode direction default, weights at zero except direction_weight 0.25). The node captures its low-grid trajectory internally, so it wants a Flow Trajectory handle but not a separate Trajectory Capture node. Outputs: patched model and metrics.

Install

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

Restart. No extra Python deps, no bundled weights - you bring the H3 model and, for realistic video use, the surrounding ecosystem patches your workflow already depends on.

Gotchas, and these are real

The schedule must be a full 1→0 H3 sigma schedule; partial low-sigma refinement schedules are rejected outright because the absolute flow origin gets ambiguous once you've trimmed the front. Sampler objects carrying an explicit external noise_sampler closure are refused - mutable RNG can't safely cross the geometry reset. And the honest caveat from the author's own evidence: the direction-only 14-step reference was the quality winner in their matched difficult-motion sweep, but this is research-grade territory with a limited prompt set, not a universal speedup guarantee. Test on your own motion before you trust it on your money shot.