Teacher Reference Caches
Ready-made teacher reference artifacts for the teacher-normalized
distribution evaluation of
DistillAlign. They replace
the teacher-sample + extract stage entirely: pass the .npz file to
--teacher-features and only the student side needs to be generated.
Contents
wan2.1_t2v_1.3b_reference_vjepa2.npz V-JEPA2 features, 256 x 2560
wan2.1_t2v_1.3b_reference_vjepa2.meta.json extraction + sampling protocol
wan2.1_t2v_14b_reference_vjepa2.npz
wan2.1_t2v_14b_reference_vjepa2.meta.json
videos/wan2.1_t2v_1.3b_reference/ 256 reference videos (p{prompt}_s{seed}.mp4)
videos/wan2.1_t2v_14b_reference/ 256 reference videos
Protocol
Both caches follow the paper protocol:
- Teacher: Wan2.1-T2V-1.3B / Wan2.1-T2V-14B, 25-step UniPC, timestep shift 8.0, classifier-free guidance 5.0, official negative prompt.
- Grid: 16 prompts (
prompts/distribution_eval_16.txt) x 16 seeds (11,22,33,42,44,55,66,77,88,123,456,789,2024,3407,7777,9999), 256 samples per teacher. - Features:
facebook/vjepa2-vith-fpc64-256, 8 uniform frames from the first 81 frames, token mean/std concat pooling, L2-normalized, 2560-D.
The exact extractor fingerprint and prompt/seed hash are embedded in each
.npz and validated by the metric CLI, which rejects incompatible student
caches instead of producing incomparable numbers.
Usage
python evaluate_distribution.py run \
--checkpoint checkpoints/distillalign_distill_14b_teacher.pt \
--jobs outputs/eval_jobs.jsonl \
--teacher-model Wan2.1-T2V-14B \
--teacher-features teacher_caches/wan2.1_t2v_14b_reference_vjepa2.npz \
--output-dir outputs/my_eval
The videos are provided for inspection and for re-extraction under future
feature encoders; they are not needed when using the .npz caches.