MiniMax H3 Motion Adapter: Real 240fps Motion for Fast Action

The problem

MAINodes' Motion Lab regenerates the fast span of a MiniMax H3 clip on a stretched clock (the "de-rope") and recovers it back to real time. On very fast motion the base model over-produces: the result alternates between advance and snap frame to frame, and the pass invents about 40 percent more movement than the shot called for (measured rate 1.416 instead of 1.0).

The motion adapter is a plain rank-16 weight delta — no extra compute, no custom node required. It was trained on exactly that task: hold out the frames inside a motion burst, keep the rest as clean context, and predict the held tokens. The extra clock still comes from the de-rope nodes; the LoRA just changes how the model behaves on that stretched clock.

Measured on four clips the adapter never saw during training (fight, anime sword ring, character swap, dialogue close-up), against the same graph without the adapter:

• advance/snap alternation dropped from 0.370 to 0.134 and frame-to-frame over-production from ~1.4× down to ~1.01× (render 63.6 s → 49.9 s on the same span);
• blind operator picks, adapter vs no adapter, same seed: 7 for, 0 against, 1 no preference;
• it transfers to the Ref2VA checkpoint, so one file serves both fl2va and ref2va graphs.

The comparison clip below shows the source plate, the base-model stretched pass, and the adapter pass side by side:

Without the adapter the fighters advance and snap frame to frame; with it (bottom-right) they keep moving at the speed the shot called for.

The temporal-expansion ladder

The follow-up release (2026-08-24) retrains the same rank-16, attention-only construction against real intermediate frames: windows cut from the public GOPRO_Large deblurring dataset (240 fps source), decimated to the working frame rate and re-expanded, so held tokens are scored against frames a real camera actually captured. The world clock in training is truthful — each token is labelled with the real time it spans.

Every 25th training step is published (temporal_expansion/step025 … step375, plus warm100, 63 MB each). Held-token error, scenes the adapter never saw (no-adapter baseline 1.23, plain interpolation 0.26):

The curve never turns, but the numbers and the eye disagree about where to stop: step 100 is the playback pick (fast motion smooth, clip still feels like itself), while step 375 measures best but can read over-smooth. On a static-content probe the family does not invent motion — it lands below the interpolation floor rather than above it.

MAINodes update

The same update adds new alpha nodes to ComfyUI-MAINodes (2026-08-24): H3 Repair Plan / Repair Splice (re-render a bad frame range, snapped to whole tokens, regenerate through the nearest shot cut, splice everything else back bit-exact), H3 DyRoPE (per-block clock presets: "timing fidelity" blocks 30–49, "minimal shimmer" 40–49), audio prefix freeze on H3 V2V Init, H3 Drift Control (chains stop washing out, joins hold 0.85–0.89 through 4 links), H3 Delta Color Carry (cancels VAE round-trip darkening), and a compare-deck builder for A/B review pages.

Use in ComfyUI

Drop a file into models/loras/minimax_h3/ and load it with a stock LoraLoaderModelOnly at strength 1.0 on the model that feeds the de-rope pass — not the first text-to-video pass. A ready graph ships in ComfyUI-MAINodes as examples/motion_pipeline_adapter_api.json.

Settings that work (from the model card):

• Strength 1.0 is the smoothness setting; 0.75 invents fewer objects.
• Inject is the dial: 0.45 for character/dialogue where the base over-produces hard; 0.30 where identity or props are the deliverable.
• Keep the tail guide on; apply the adapter to the de-rope pass only.
• Rough edges: it costs about 1 dB of anchor fidelity on native keyframes, over-corrects calm chains (keep it off on calm content), and mutes strong colour / thins particle effects — worst on Ref2VA and full-clip passes.

The adapter is MIT-licensed; using it with the base model is subject to the MiniMax H3 Community License Agreement.

Links

• Hugging Face: MATLOWAI/MiniMax-H3-Motion-Adapter
• GitHub: matlowai/ComfyUI-MAINodes