noctuashap/trackwan-1.3b-bidir-merged-bias

TrackWan 1.3B — bidirectional I2V + point-track (merged-bias, ckpt-4800)

A track-conditioned finetune of Wan2.1-Fun-1.3B (InP) for image-to-video generation with MotionStream-style point-track motion control. This is the merged_bias bidirectional training run, checkpoint 4800.

• Base model: Wan-AI/Wan2.1-Fun-1.3B-InP (Apache-2.0)
• What's trained here: the WanTransformer3DModel with the added track-conditioning head (transformer/model.safetensors, ~2.9 GB). Text encoder / VAE are the standard Wan components.
• Type: research checkpoint.

⚠️ Requires FastVideo (custom architecture)

This is not a vanilla diffusers pipeline — the track-conditioning is a custom WanTrack architecture. Load it with the FastVideo fork that produced it, e.g.:

# in the FastVideo repo, using the trackwan inference helper
import data_pipeline.trackwan_infer as twi
model, cfg = twi.load_trackwan("<this repo checked out>", "<train yaml>")

Pair it with the base Wan2.1-Fun-1.3B-InP text encoder + VAE. See the FastVideo project for the exact config/YAML and track-conditioning input format (per-frame point tracks + visibility).

Files

• model.safetensors — trained transformer weights (bf16).
• config.json — transformer architecture config.

License

Apache-2.0, inherited from the Wan base model.