MiniMax H3 Progressive Handoff (Target Input)

This is the pack's flagship: the progressive handoff that works inside Continuum. H3 at ~1 MP is slow enough that shaving 20% of the wall time is the difference between "let it render while I make tea" and "leave it overnight," and this is the node that demonstrably does it - the author's measured run went from ~777 s for a proper two-pass upscale/refine to ~622 s for a progressive learned-handoff run at a comparable final size, with more final pixels. That's ~1.25× observed speedup. Not a controlled same-seed benchmark, they'll be the first to tell you, but a real workflow-level number.

The trick is a private grid. Continuum's session, native masks, and bookkeeping have to stay target-sized, so this node cheats: it runs the early part of the schedule on a smaller private video state it owns internally, then transfers to the real target grid at the handoff and starts a fresh sampler lifetime there. The rest of the graph never sees a geometry change. Audio never gets spatially touched. All the history-reset hygiene from the plain progressive node applies - exact low-grid probe at the boundary, Spectrum feature history cleared, first high-grid call forced actual.

Wiring (this is the one to get right):

DiffAid -> Untwist -> Spectrum -> Progressive Handoff (Target Input) -> Continuum

Create one Flow Trajectory handle and connect it here - the node captures the low-grid trajectory internally, so no separate Trajectory Capture node on this path.

Inputs that matter. source_mode of pixels (explicit source_width/source_height, e.g. 864×640) or scale, where source_scale is a linear H/W scale, not a megapixel fraction: private area ≈ target × scale². At ~1 MP, source_scale=0.70 was the author's tested quality/compute sweet spot (~832×640 → 1184×896) while 0.65 started losing reference likeness. Keep handoff_coordinate at 0.35, handoff_selection fixed, guidance_mode direction with direction_weight 0.25 - this is the tested operating point, not a universal optimum. Then handoff_transfer: bicubic is the compatibility default, and learned_3d is the quality upgrade - it routes the exact-probe clean video state through a connected H3_LATENT_UPSCALER provider from the companion repo (xmarre/Comfyui_Minimax_h3_latent_Upscaler) instead of resizing it, which visibly fixed bicubic's handoff artifacts at aggressive source scales. One extra CNN inference per chunk, zero added H3 NFE.

Install

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

Restart. For the learned transfer you additionally install the companion upscaler repo and create its provider node. No bundled models either way.

Honest caveats

Full 1→0 sigma schedule required; partial refine schedules are rejected. Private-grid noise is a standard-Gaussian CPU generator keyed to your graph seed, so exotic custom noise samplers won't survive the geometry reset. And the ~20% figure is against one specific proper 7+6 two-pass workflow - great evidence, not a law of physics. The good news: decoded media, not just telemetry, backs the quality claims at the tested operating point. Start there, then experiment.