Repository navigation
3.35.0
Highlights
Text-to-motion v8.0 — generated animations now face the right way (#837, PR #963)
qtmesh anim <rig> --generate "<prompt>" --model produced motion that moved convincingly but oriented backwards, and its walk had a defect that read as a flat knee with a false second knee below it. Both turned out to be problems with the training data, not the model architecture.
-
Trained on the curated template clips instead of the CMU corpus. Measured against a real reference walk, CMU "walk" windows sit 1.73–2.07 rad away — further than a real punch, dance or jump, against a real-vs-real floor of 0.437. The corpus never carried the correct facing convention, so the model could not learn it: v6.x scored 1.89 while its own training data scored 1.73. The curated clips measure 0.303–0.462, and the model now trains on those.
-
The walk's bend is back in the knee. The old model bent the knee 25° and the ankle 31°; a real walk is 38°/10°. A near-straight knee with the stride absorbed by the ankle is what rendered as a second joint. A new leg-chain loss targets exactly that ratio.
v7.x v8.0 real reference knee / ankle 25° / 31° 41° / 8° 38° / 10° leg-chain penalty 0.359 0.0003 0.014 walk refDist 0.287 0.243 0.437 (floor) run refDist 0.472 0.324 — -
Four broken actions removed. An anatomical ankle gate (rejecting clips whose foot folds past 70° from the shin) and a minimum-resolved-roles floor caught
march(106°),throw(101°),hey(90°) andconfession— all built on source clips where the foot sat perpendicular to the shin in 100% of windows, or where the rig resolved too few joints for the model to have any leg data at all. Those prompts now fall back to the curated template library, which formarch/throwis strictly better than what the model produced.march/marchingresolves towalk. -
Measured across 20 actions, v8.0 is closer to ground truth on 8, further on 3, and tied on 9. The largest gain is
wave(0.710 → 0.082) — the arm now actually goes overhead instead of hanging at the side.
The model downloads on first use; no action needed. The previous model is preserved on HuggingFace for rollback, and the 150-clip library the model was trained on is published alongside it.
Full changelog: 3.34.0...3.35.0