georgeNakayama/lod_generation

Wan 2.1 Vchitect Option-1 training cache

This gated dataset contains 100,000 aligned training examples used by the Wan Option-1 run in foveated_diffusion_3d:

• original RGB source videos;
• FP16 Wan VAE latents at full resolution only, with shape (1, 16, 21, 60, 104);
• prompts and sample/video mappings in manifest.jsonl; and
• the 480x832x81 Procrustes artifact used by the active training configuration.

The source cache held 16 anisotropic pyramid resolutions. The active loader only reads payload["latents_pyramid"][(1, 1)], so the other 15 resolutions were intentionally omitted. Each rewritten .pt retains the original payload shape and metadata but its latents_pyramid dictionary contains only (1, 1).

Layout

wan21_vchitect_81f_480x832/
  latents_hr_fp16/shard_00000.tar ...
  rgb_videos/shard_00000.tar ...
  manifest.jsonl
  metadata.json
  meta.pt
  shards.jsonl
  artifacts/procrustes_apost_gridrange_subnorm_centercrop_480x832_81f.pt

Shard pairs use the same 1,000 selected indices. The manifest records both tar paths and member names for every sample. RGB files preserve their original bytes and basenames. The original absolute video_path is retained for provenance; after extraction it must be rewritten to the local video directory.

Raw full-resolution tensor payload: 419.328 GB. RGB videos: 368.591 GB. Tar/serialization overhead is small.

Access and use of the RGB videos remain subject to the terms of their source dataset. Wan model weights and the training source tree are not included.