JiaqiFeng/Temporal-NoPE

Temporal NoPE runtime checkpoint

This public repository carries the immutable stable causal-video checkpoint used as the common initialization for the Temporal NoPE experiments in Aurora-edu/NoPE.

Checkpoint identity

• File: checkpoints/framewise/causal_cd.pt
• Size: 5,676,220,819 bytes
• SHA-256: c951a6b4804cd637fecfc857e9a59d54b6e4cd7846c38360baa3cf3b525f0d22
• State-dict key: generator_ema
• Architecture: Wan 2.1 T2V 1.3B causal model

The portable configuration constructs the exact 30-layer, 1536-dimensional, 12-head architecture without first downloading the redundant public base-DiT weights, then strictly loads all 825 checkpoint entries. The strict load gate passes with no missing or unexpected keys.

Download

hf download JiaqiFeng/Temporal-NoPE \
  checkpoints/framewise/causal_cd.pt \
  configs/temporal_nope/d4_source_rope_uniformscale_c010_portable.yaml \
  wan_models/Wan2.1-T2V-1.3B/Wan2.1_VAE.pth \
  wan_models/Wan2.1-T2V-1.3B/config.json \
  prompt_cache/eval_umt5_bf16_lmdb/data.mdb \
  --local-dir /path/to/NoPE

The bundle includes the Wan VAE and the read-only eval prompt-embedding LMDB, so cached-prompt visual preflights do not need to load the UMT5 encoder. The public base DiT is intentionally omitted because the portable config strictly loads every model parameter from causal_cd.pt.

Always verify the SHA-256 checksum before running an experiment.

Recovery-v3 visual evidence

The repository also preserves the complete small-gate outputs for the final attention-scaling diagnosis under outputs/temporal_nope/:

• recovery_v3_d4_source_rope_uniformscale_c010_portable_micro_visual_gate_seed0_illidan
• recovery_v3_d5_source_rope_uniformscale_c030_portable_micro_visual_gate_seed0_illidan
• recovery_v3_attndiag_a0_c000_val004_seed0_illidan
• recovery_v3_attndiag_d4_c010_val004_seed0_illidan

These 44 files include all generated MP4s, resolved configs, sample metadata, rank completion markers, and raw temporal-attention records. Both D4 and D5 failed mandatory human review; the outputs are failure evidence, not promoted model samples. The GitHub reports contain the formal visual decisions and mechanism analysis.