3D-WM Atomic — v3 (batches b07–b14)
Synthetic CAD construction videos paired with per-frame atomic action ops, for training video→action (inverse-dynamics) and image→video world models.
v3 extends the v2 dataset (batches b01–b06) with b07–b14. Identical format and contract.
Format: v2-continuous-mm
Raw float-millimeter op parameters (no quantization); each clip is translation-canonicalized so the final-mesh bbox is centered at the world origin. The full schema — atomic op vocabulary, scale/camera metadata, the SKETCH_BEGIN depth-DOF caveat, and the recommended near-blank reader filter — is documented in the companion dataset_spec.md.
Layout
data_hula_b07 … data_hula_b14/ # one prefix per source batch
b0000 … bNNNN/train/shard-*.tar # tar shards, ~25 clips each (~10 shards/batch-idx)
Each clip = one (case, view):
0000.png … NNNN.png— per-op construction-animation frames (256×256)meta.json— scale / camera / canonicalization + per-frameop_spansops.json— atomic op tuples (raw-mm floats) + human-readableops_human
5 views per case (iso / front / side / top / random) sharing one auto-fit scale anchor — identical world_diag_mm / approx_mm_per_pixel across siblings, only the camera angle differs.
Quality gates (at generation time)
- Geometry validity — OCC
BRepCheck; invalid / self-intersecting solids are dropped. - Near-blank — a case that renders near-blank from every view (measured on a neutral palette) is dropped (valid-but-tiny solids).
Every shipped clip is therefore a valid, visibly non-blank solid.
Scale
~8000 cases/batch × 8 batches × 5 views ≈ 300k clips.
🤖 Generated with Claude Code