MiniMax H3 Flow-Aligned Refine State

If you run H3 through Continuum's integrated learned upscale/refine, Flow-Aligned Regenerate is aimed at the wrong target. Continuum hands you a H3_CONTINUUM_REFINE_STATE object for its per-chunk refinement, not a bare model - so this node exists to patch that, carrying the same flow-aligned guidance into the integrated refine pass.

Think of it as the Continuum-flavored twin of Regenerate: same guidance menu, same philosophy, different socket. It takes the refine_state Continuum emits, finds the H3 model and positive conditioning tucked inside it, applies the trajectory guidance to the model, and returns a patched refine_state you can hand to the companion learned upscaler/refiner. You wire one of these per chunk's refine state.

The inputs that matter. refine_state is Continuum's output; trajectory is your shared Flow Trajectory handle from pass one. guidance_mode defaults to direction, and like its sibling the conservative recommendation is to keep it there - direction+acceleration, direction+temporal, and downsample_consistency are implemented and validated but have no demonstrated decoded-media advantage in the author's matched tests. Set direction_weight, acceleration_weight, consistency_weight, temporal_weight accordingly; zero is the right value for anything you're not actively researching. Outputs are the patched refine_state and a metrics object you can chain into a shared artifact.

This node also quietly validates that guidance is safe to apply: it checks the state is Continuum refine-state API 1, confirms the model carries a captured trajectory run id, and confirms the conditioning provenance matches. No match, no guidance - it raises instead of guessing, because applying a trajectory from a different run to a refined chunk would be actively harmful rather than merely useless.

Install

Same pack, same one-liner:

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

Restart ComfyUI. The real prerequisite is a working Continuum integrated-refine pipeline with the companion H3 latent upscaler/refiner - this pack's nodes are the middleware, not the whole meal.

The ordering rule that bites everyone

Trajectory Capture must be the final model patch before Continuum, so the refine state carries trajectory provenance. If it isn't, Flow-Aligned Refine State will refuse the state with a message about missing captured provenance - and the fix is reordering your patch chain, not retuning weights. Also remember Continuum keeps running its learned upscaler/refiner; this node only changes how the refine pass is guided toward the low-res trajectory, so it slots between Continuum's first pass and the refine, per chunk, with one shared handle across them all. Keep audio locked from pass one; H3's audio track isn't something this guidance is allowed to invent fresh.