MiniMax H3 Trajectory Capture

If you've ever done a hires-fix-style workflow - generate small, then upscale and re-run - you've felt the waste: pass two re-samples most of the same trajectory it already walked once. MiniMax H3 Trajectory Capture is the node that records pass one's footsteps so the rest of this pack can stop the second pass from repeating them.

It patches an H3 model so that, while the sampler runs, every predicted-clean estimate (the model's x0 guess at each denoising point) is written into a H3_FLOW_TRAJECTORY handle instead of vanishing into the next step. That trajectory is the raw material for Flow-Aligned Regenerate's low-frequency guidance on a later high-res pass, or for the progressive handoff nodes that grow the grid mid-schedule. It changes nothing about the actual sampling - the recorded model produces identical output - it just has a memory now.

Inputs you actually touch: model is your low-resolution H3 model patch, trajectory is the handle from a MiniMax H3 Flow Trajectory node (create one and reuse it), and capture_forecasts defaults to False for a good reason. When False, only exact H3 evaluations become trajectory anchors. When True, Spectrum's cheap forecast calls are captured too - but they're provenance-tagged and never treated as trustworthy anchors, because a forecast is a prediction of what the transformer would say, not what it said. Turn it on only if you're researching forecast behavior. For quality work, leave it off.

Outputs: a patched model (wire it into your first-pass sampler instead of the raw one) and a metrics object you can chain forward. Feed that metrics into the downstream Flow-Aligned Regenerate node's optional metrics input when you want one combined benchmark artifact for the whole two-pass run, rather than a separate blob per stage.

The two-pass recipe it belongs to

1. Create one Flow Trajectory handle.
2. Patch your low-res H3 model with Trajectory Capture and run the base sample.
3. Do the existing learned latent upscale / second-pass initialization.
4. Patch the high-res model with Flow-Aligned Regenerate (or each Continuum refine_state with Flow-Aligned Refine State).
5. Same trajectory handle everywhere. Keep pass-one audio locked in the surrounding workflow.

Install

From ComfyUI/custom_nodes:

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

Restart ComfyUI. No extra Python packages, no bundled models. The real prerequisite is the wider H3 ecosystem this pack assumes you already run: native ComfyUI H3 support, and for the Continuum path, Continuum plus whatever upstream patches (Spectrum, DiffAid, untwisting RoPE) your workflow already has.

Gotchas

This is the one people trip on with the integrated learned-refine path: Trajectory Capture must be the final model patch before Continuum, so the emitted refine_state carries trajectory provenance. Put it earlier in the chain and Flow-Aligned Refine State will refuse the state outright. And because the capture happens during an actual sampler run, you can't record a trajectory that was generated last week - the capture, the guidance, and the shared handle all have to live in the same graph execution.