Downloads · 30 days
0
cad-video-gen-exp1
hz6666/cad-video-gen-exp1
cad-video-gen-exp1 is a machine learning model from hz6666. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Video-gen checkpoints (DiT + VACE) for the CAD world model.
Access
Public
Updated Jun 16, 2026
Repo size
189 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors129 GB · 100%
At a glance
- Access
- Public
- Created
- Jun 4, 2026
- Updated
- Jun 16, 2026
- SHA
- 1ce0f360
Model card
From the Hugging Face model README
cad-video-gen-exp1
Video-gen checkpoints (DiT + VACE) for the CAD world model.
25f xycanon line (2026-06-13)
Reverse-DF 25-frame video base + op-decoders on the xycanon data (canonical 5-view
renders). Full live asset map: cad_video_gen/docs/data_inventory.md.
| file | what | metric |
|---|---|---|
exp2_df_xy25_rev/step-16670.safetensors | 25f xycanon video base (reverse-DF, C2 target@frame0); the frozen backbone every op-decoder reads for spatial memory | video LPIPS 0.298 (vs 17f base 0.191, same 24 cases) |
exp2_opdec_xy25_mv/opdec-mv-step-8000.pt | multi-view op-decoder (5 fused views -> one op program) | exec 30%, exact 0.125, len_mae 2.09 |
exp2_opdec_xy25/opdec-step-8000.pt | single-view op-decoder (baseline) | exec 9%, exact 0.031, len_mae 6.11 |
labels/labels_fullprog_xycanon25.pt | 25f xycanon op programs (op_type/cont/cat) | op-decoder targets |
labels/labels_xycanon25_full.pt | uids in cache row-order | — |
labels/labels_xycanon25.pt | held-out subset | — |
Build-IoU still ~0 — op structure is right (multi-view lifts exec 9%->30%) but geometry/coord precision is gated by this base; the v2+v3+xy stronger no-view base is the next iteration.
- Created
- Jun 4, 2026
- Updated
- Jun 16, 2026