dancing-stick-figures-baselines

Dancing Stick Figures — reference models

Reference checkpoints for Dancing Stick Figures, the dataset introduced in Dancing Stick Figures: An Introductory Dataset for Training Video Generation Models. The current paper release is under paper-v6/. Historical image, short-video, and autoregressive baselines remain at the repository root.

Start here: Dataset · Colab · Code

Paper v6 checkpoints

All learned video rows below generate text-conditioned 64-frame, 64×64 clips at 20 fps. The Pixel DiTs operate directly on RGBA pixels. Mini-Wan generates normalized video latents and must be used with the released f4t4d8 codec and latent statistics.

The Pixel video runs use batch 8 × accumulation 2, velocity-prediction flow matching, foreground-weighted pixel loss, and an RGBA clean-prediction auxiliary term of weight 1. Mini-Wan uses a latent flow objective plus its decoded-RGBA auxiliary. These auxiliary constructions are analogous but not mathematically identical.

Exact hashes, architecture fields, provenance, and evaluation-file mappings are in paper-v6/release_manifest.json. Canonical evaluation JSONs are under paper-v6/evaluations/, and fixed-seed 64-frame samples are under paper-v6/samples/.

Table 3 reference results

The video protocol uses 128 held-out source motions, sampling seeds 0/1/2, 64 frames at stride 1, 50 Euler steps, and CFG 3. Lower is better for TVR, LIE, CPE, jerk, and FVD. Speed and motion fraction are two-sided diagnostic signals to compare with the real-reference row.

Download the current release

hf download sprited/dancing-stick-figures-baselines \
  --include "paper-v6/*" --local-dir checkpoints

Checkpoint dictionaries contain EMA model weights, the optimizer step, and saved architecture arguments. The codec checkpoint additionally contains its model state and training metadata. Use the matching trainer/evaluator revision from the linked code repository; these are research checkpoints rather than a packaged inference API.

Historical baselines

The repository root retains the completed v0.1 image, 8-frame UNet, and long autoregressive UNet references for reproducing earlier dataset-card results. Files explicitly labelled interim, superseded 10k Mini-Wan checkpoints, and the no-decode-loss Mini-Wan checkpoint are not part of the current release.

Data are CC0-1.0 and code/checkpoint wrappers are MIT. See the dataset card and paper for scope and limitations.