Track-v0
Track-v0 is a 1.3B causal WanTrack image-to-video checkpoint trained with sparse point trajectories.
Training recipe
- Initialization: WanTrack bidirectional stage-2 checkpoint 600
- Teacher Forcing: 3,000 steps
- Causal Distillation: 2,000 steps
- Self Forcing: 1,000 steps
- Published weights: Self Forcing checkpoint-1000
student_ema - Attention: 3 latent frames per block, local attention 6, sink size 1, relativistic RoPE
- Training resolution: 832 x 480, 121 pixel frames / 31 latent frames
This repository contains the complete model bundle plus the full EMA transformer. The frozen TrackEncoder parameters are preserved from the initialization checkpoint; the remaining transformer parameters come from the final SF EMA.
Use this checkpoint with the FastVideo WanTrack causal model integration. See training_config.yaml for the exact model and pipeline configuration and export_receipt.json for provenance hashes.
Integrity
- SF EMA SHA256:
fcc60e509917996af05e598ca1c58b4ef6368318b55eaf2f9879d9b031a9befd - Full transformer SHA256:
3daa8dfa39ea4623415f6c128830058c8ae84c4a01c2144d3631d81244a4c959 - FastVideo source commit:
3a398a2bbd90ce511aabbb39996f1d898862192f