README.md

metadata

license: apache-2.0
task_categories:
  - text-to-video
  - image-to-video
tags:
  - wantrack
  - trackwan
  - motion-conditioned
  - point-tracks
  - synthetic
  - fastvideo

WanTrack synth toy — 720p

A 720p/24fps recreation of noctuashap/wantrack_synth_toy: one shared seed image + 50 motion captions → 50 I2V clips, each varying only the motion. Used as the overfit set for the bidirectional TrackWan teacher recipe in FastVideo.

How it was built

1. Generation — Wan2.1-I2V-14B-720P, conditioned on the single synthetic_seed.png + each line of captions.txt, 720×1280, 121 frames @ 24fps (gen_synth_i2v_worker.py).
2. Tracks — CoTracker3 on a 50×50 grid + SAM object segmentation (extract_tracks.py --segment).
3. Preprocess (Stage 5) — VAE latents + T5 text embeds + CLIP frame-0 + track points/visibility baked into parquet (v1_preprocess --preprocess_task i2v_track, train_fps 24).

Layout

path	what
videos/vid_000000..000049.mp4	50 generated clips, 720×1280 @ 24fps, 121 frames
tracks/*.npz	CoTracker+SAM tracks (points, visibility, object_ids, weights)
videos2caption.json	manifest: caption, fps, duration, resolution, points_path
preprocessed_i2v_track/combined_parquet_dataset/	training-ready parquet (1 file, 50 rows)
synthetic_seed.png	the shared 480p seed frame (I2V resizes to 720p)
captions.txt	the 50 source motion captions

Parquet schema (per row, pyarrow_schema_i2v_track)

vae_latent/first_frame_latent [16, 31, 90, 160] (float32) · clip_feature [257, 1280] · track_points [121, 2500, 2] · track_visibility [121, 2500] · object_ids, track_weights · text_embedding. Point the trainer's data_path at preprocessed_i2v_track/combined_parquet_dataset.

Notes

• ~1 in 5 captions involve a human hand manipulating objects — carried over from the original toy set (a deliberate "external agent" motion category), not a generation artifact.
• All clips share one seed frame, so appearance is fixed and only motion varies — ideal for overfitting/validating a point-track conditioning pathway.