Dataset Viewer (First 5GB)
Auto-converted to Parquet Duplicate
pth
listlengths
3
3
__key__
stringlengths
17
21
__url__
stringclasses
1 value
[{"context":null,"input_latents":[[[[[0.6015625,0.74609375,0.71484375,0.671875,0.69921875,0.6171875,(...TRUNCATED)
cache_v23_rev/0/11547
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
[{"context":null,"input_latents":[[[[[0.59375,0.73828125,0.71484375,0.6328125,0.578125,0.58203125,0.(...TRUNCATED)
cache_v23_rev/0/4292
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
[{"context":null,"input_latents":[[[[[0.59765625,0.74609375,0.7109375,0.6484375,0.69140625,0.6210937(...TRUNCATED)
cache_v23_rev/0/3834
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
[{"context":null,"input_latents":[[[[[0.60546875,0.73828125,0.6953125,0.59375,0.60546875,0.67578125,(...TRUNCATED)
cache_v23_rev/0/3963
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
[{"context":null,"input_latents":[[[[[0.59765625,0.74609375,0.72265625,0.66796875,0.6953125,0.597656(...TRUNCATED)
cache_v23_rev/0/6795
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
[{"context":null,"input_latents":[[[[[0.59765625,0.74609375,0.7109375,0.6484375,0.6953125,0.62109375(...TRUNCATED)
cache_v23_rev/0/3900
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
[{"context":null,"input_latents":[[[[[0.6015625,0.74609375,0.69921875,0.67578125,0.6875,0.5859375,0.(...TRUNCATED)
cache_v23_rev/0/3074
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
[{"context":null,"input_latents":[[[[[0.59375,0.73828125,0.71484375,0.63671875,0.58203125,0.56640625(...TRUNCATED)
cache_v23_rev/0/1674
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
[{"context":null,"input_latents":[[[[[0.60546875,0.734375,0.734375,0.671875,0.7109375,0.6171875,0.72(...TRUNCATED)
cache_v23_rev/0/5322
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
[{"context":null,"input_latents":[[[[[0.59375,0.74609375,0.72265625,0.67578125,0.703125,0.62109375,0(...TRUNCATED)
cache_v23_rev/0/3549
hf://datasets/hz6666/cad-video-cache@b193f3f33f1c7f909980095b50855f74be779272/cache_v23_rev.tar
End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

cad-video-cache — pre-encoded latent caches (skip data_process)

Slimmed VAE+T5 latent caches for the CAD video world-model (cad_video_gen). Download + extract a tar to skip the ~8-9h data_process, then train directly: train_df.py --dataset_base_path cache_xxx.

What's in a .pth

Each clip is one file cache_*/0/{idx}.pth (~1.3 MB), a 2-tuple of dicts:

  • [0]["input_latents"] — Wan-VAE video latent (1, 16, T, 60, 60) bf16: 16 channels, T temporal-latent frames (VAE 4:1 causal), 60x60 = 480/8 spatial. T=7 -> 25 pixel frames, T=5 -> 17 pixel frames.
  • [1]["context"] — T5-XXL text embedding (1, 64, 4096) (slim keeps the first 64 tokens).

⚠️ recipe-bound: a cache is reusable ONLY for the same (frames, resolution, reverse, slim) recipe it was built with — a different num_frames / resolution needs a fresh data_process.

Caches & their provenance

tar status pth frames source dataset pipeline use
cache_xycanon25_rev.tar available 16670 25f (latent T=7) xycanon — canonical 5-view (3334 parts x 5 views) xycanon clips -> make_vace_dataset (atomic_xycanon_vace25) -> reverse_videos -> data_process 25f flagship exp2_df_xy25_rev (LPIPS 0.298) + the frozen-backbone memory for all op-decoders
cache_v23_rev.tar available 11748 17f (latent T=5) v2+v3 — curated 11748-uid set, iso-view (v00) fetch v00 clips -> make_vace -> reverse -> data_process 17f base line / legacy op-decoder
cache_xycanon_rev.tar available 10201 17f (latent T=5) xycanon — earlier subset (pre-25f line) xycanon clips -> make_vace -> reverse -> data_process early 17f xycanon experiments
cache_v2v3xy25_rev.tar available ~34938 25f (latent T=7) v2+v3+xy merged, no-view prompts (18268 v23 + 16670 xy) merge symlinks -> reverse -> data_process the new no-view v2v3+xy video base

Common to all: 480x480, reverse (target rendered at frame 0), slim (only input_latents + context[:64]), bf16. Raw sources live in hz6666/3d-wm-atomic-{v2,v3} (WebDataset shards) and the xycanon repo; the rebuild recipe is data_backup.md in cad_video_gen.

Extract: tar -xf cache_xycanon25_rev.tar -> cache_xycanon25_rev/. Full live asset map + lineage diagram: docs/data_inventory.md in the cad_video_gen repo.

Downloads last month
42