trackwan-14b-720p-overfit-warmstart

TrackWan 14B / 720p — Overfit warmstart checkpoint (Stage B, step 3000)

Training-state checkpoint (PyTorch DCP) from Phase 0 / Stage B of the bidirectional TrackWan teacher recipe (point-track-conditioned I2V, 14B, 720p), built in FastVideo.

What this is

• Stage B overfit of the track pathway (track_encoder + patch-embed track slot), head trainable, on a small 50-clip 720p synthetic set. Random track-ID sampling (WANTRACK_FIXED_SAMPLE=0) so the pathway generalizes over arbitrary ID/point assignments (the invariance stage — it can only be learned while the head is trainable, before stage-1 freezes it).
• Warm-started from Stage A (checkpoint-2000): fixed-ID overfit → random-ID overfit, continued 2000→3000.
• Maskless and CLIP-free by design: WANTRACK_PMASK=0 (mid-frame masking belongs to stage-2) and WANTRACK_IMAGE_COND=0 (image cross-attention is enabled from stage-1 on the pristine merged base; turning it on mid-overfit shocks a CLIP-free-trained seed).
• Config: lr 1e-4 (constant), flow_shift 6, sparse conditioning (WANTRACK_SPARSE=1, EXTRA_RANDOM=20), d64 track-ID embedding + bias track encoder.

Format & caveats

• Full training-state checkpoint: model + optimizer + scheduler + per-rank RNG, sharded as dcp/__<rank>_0.distcp (8 ranks, HSDP 2×4). Not a from_pretrained model — consolidate via the FastVideo export step (03_export.sh) to get a diffusers transformer.
• The base is discarded by the Phase-1 merge — only the (now ID/subset-invariant) track pathway is lifted onto a pristine Wan2.1-I2V-14B base. So this checkpoint's value is the track pathway; it is primarily a reproducibility / re-merge-source artifact.

Provenance

Stage A (fixed IDs) → Stage B (random IDs, this ckpt) → export → merge → OpenVid stage-1 (CLIP on) → synthetic stage-2 (masking). See the FastVideo data_pipeline/720_stage_1/ recipe.