When De-Rope Stretches the Clock

MiniMax-H3’s Motion Lab in ComfyUI-MAINodes already has a clever trick for fast motion: regenerate the hot span on a stretched clock (the de-rope), then fold it back to real time. The catch? On very fast bursts the base model loves to over-produce — fighters advance, snap, advance, snap — and your timeline starts stuttering like a bad teleport.

MATLOWAI just shipped a pilot fix you can drop straight onto that pass: a rank-16 LoRA trained for the de-rope situation itself, plus a newer temporal-expansion ladder warm-started from the same recipe against real intermediate frames.

Not a General-Purpose Motion LoRA

This file is not “sprinkle motion on any H3 render.” Load it on a plain first-pass text-to-video and you mostly buy the costs (colour muting, over-smoothing) with none of the benefit.

Wire it like this:

  1. Put minimax_h3_motion_adapter_pilot_r16.safetensors (63 MB) in models/loras/minimax_h3/
  2. Attach a stock LoraLoaderModelOnly to the model that feeds the de-rope / Motion Lab pass only
  3. Skip the first T2V pass — leave that one alone
  4. Grab the shipped example: examples/motion_pipeline_adapter_api.json (API format; also lives on the Hub)

It transfers to Ref2VA as well as fl2va, so one adapter serves both graph families.

Pins First, Adapter Second

MATLOWAI’s exhibits are delightfully honest about who does the heavy lifting.

Pinning the window’s first and last frames to the plate frames they must butt against stops most splice jumps for free — two extra wires, no extra wall time in their measured tile. The adapter then spends the spare stretched-clock capacity on smoothness on top of that.

So if your seams are jumping: pin first. Reach for the LoRA after the boundaries behave.

What It Actually Wins

Trained as a pilot (rank 16, a few hours, 38 self-generated H3 clips, holdout-infilling inside motion bursts), then measured on four never-seen clips (fight, anime sword ring, character swap, dialogue close-up):

  • Advance/snap alternation improved on 4 of 4 (roughly 4.6–6.1× the noise floor)
  • Frame-to-frame over-production dropped from about 1.76–2.93 down to 1.09–1.23
  • Blind operator picks (adapter vs no adapter, same seed): 7 for, 0 against, 1 no preference
  • Smoke test on a never-seen neon-rooftop fight remix kept plate signage the no-adapter pass rewrote, with smoother alternation/jitter

Rate near 1.0 means motion lands at the speed the shot asked for; rates like 1.4 mean the pass invented ~40% more movement than it should have. The adapter’s job is teaching the model to spend extra clock on smoothness instead of invention.

Tradeoffs (Keep It Honest)

Creators deserve the downside list in plain language:

  • About ~1 dB of anchor fidelity cost on native keyframes
  • Can over-smooth calm chains — leave it off when nothing is bursting
  • Mutes strong colour and thins particles (prismatic subjects can flatten; worst on Ref2VA and full-clip de-ropes)
  • Anime-heavy wins are clearest; live-action / identity-critical work wants the lower inject values below

Settings That Work

DialStarting point

Strength

1.0 for max smoothness (won alternation in every paired cell they ran); 0.75 when you want fewer invented objects

Inject

~0.45 for character/dialogue where the base over-produces hard; ~0.30 when identity or props are the deliverable

Tail guide

Keep it on — turning it off costs identity ~3× the floor across seeds/checkpoints

Strength and inject behave like the same dial family: crank smoothness or dial back invention, then A/B on your plate.

The Temporal-Expansion Ladder (2026-08-24)

Same repo, next chapter. The pilot learned from self-generated targets; this ladder warm-starts from the pilot and retrains against real intermediate frames cut from GOPRO_Large (decimate → re-expand → score held tokens against camera truth).

Every 25th step through 375 is published under temporal_expansion/ (each still ~63 MB, same LoraLoaderModelOnly drop-in). Measured held-token error keeps falling (no-adapter baseline 1.23 → step375 0.086), but eyes and numbers disagree on the stop:

  • warm100 — the playback pick: smooth fast motion that still feels like itself
  • step375 — best on the ruler, can read over-smooth / graded in side-by-sides

If you only grab one ladder file, grab minimax_h3_temporal_expansion_warm100_r16.safetensors. Still de-rope / windowed-insertion only — not the first T2V pass.

Licence Quick Card

  • Adapter weights + graphs: MIT
  • Using them with MiniMax-H3: still subject to the MiniMax H3 Community License (shipped as LICENSE-MiniMax-H3 on the Hub)

Try It Tonight

Drop the pilot (or warm100) into models/loras/minimax_h3/, pin your window boundaries, load LoraLoaderModelOnly on pass-2 only, and start around strength 1.0 / inject 0.45 — then nudge toward 0.75 / 0.30 if invention creeps up. Compare against the demo tiles on the MAINodes adapter page before you trust a production plate.

Try Gen

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=h3-motion-adapter-derope-pilot

—Aurelia ♡