fernandotonon commited on
Commit
fb4d1c4
·
verified ·
1 Parent(s): 2a45283

docs: t2m v8.0 — flow-matching DiT, template-taught, 30 actions (#837)

Browse files
Files changed (1) hide show
  1. README.md +11 -5
README.md CHANGED
@@ -89,11 +89,17 @@ scratch** on the permissive
89
  [CMU MoCap database](http://mocap.cs.cmu.edu); beats slerp by >2× on held-out
90
  CMU motion. License: CC-BY-4.0.
91
 
92
- ## Text-to-motion — trained by us (experimental)
93
-
94
- `motion/t2m.onnx` + `motion/t2m-vocab.json` — CVAE transformer, text keyword →
95
- 22-joint world-frame clip; `motion/motion-library.json` — the curated CMU
96
- template-clip library that is the shipped default. License: CC0-1.0.
 
 
 
 
 
 
97
 
98
  ## Mesh part segmentation — trained by us
99
 
 
89
  [CMU MoCap database](http://mocap.cs.cmu.edu); beats slerp by >2× on held-out
90
  CMU motion. License: CC-BY-4.0.
91
 
92
+ ## Text-to-motion — trained by us
93
+
94
+ `motion/t2m.onnx` + `motion/t2m-vocab.json` — **v8.0**, a flow-matching DiT
95
+ (21.7M params) mapping a text keyword to a 22-joint world-frame clip, over a
96
+ 30-action vocab. Trained on the curated template clips in
97
+ `motion/motion-library-v2.json` (which is also the fallback for prompts outside
98
+ the vocab). v8.0 fixed the backwards-facing problem inherited from the CMU
99
+ corpus and a walk defect where the bend sat in the ankle rather than the knee;
100
+ `motion/t2m-v61.onnx` is kept for rollback. Full details and metrics:
101
+ [QtMeshEditor-t2m](https://huggingface.co/fernandotonon/QtMeshEditor-t2m).
102
+ License: CC0-1.0.
103
 
104
  ## Mesh part segmentation — trained by us
105