Dancing Chibi Figures — baseline checkpoints
Reference checkpoints for the sprited/dancing-chibi-figures
dataset, trained with the pure-PyTorch trainer in the dataset's GitHub repo
(train/video_ddpm.py — factorised 3D UNet, v-prediction, cosine schedule, EMA).
Why this repo exists: the dataset ships two learning tracks. The diffusers track (links at the bottom) teaches the standard tooling everyone uses; this track is the ~600-line from-scratch version where you can read every line of the model, the schedule and the sampler — and it is where the video models live (diffusers has no tiny video pipeline). These checkpoints are the "fully-trained" reference the Colab tutorial compares your short training runs against.
| file | what | steps |
|---|---|---|
unet_img64.pt |
unconditional 64px image model (--frames 1) — samples |
30k |
unet_img64_cond.pt |
group-conditional image model (--cond group, classifier-free guidance) — samples |
30k |
unet_t2v64.pt |
text-conditioned autoregressive video (--cond text --ar_ctx 8 --frames 8 --stride 2, CLIP prompt embeddings, warm-started from unet_img64) — landing soon |
60k |
unet_t2v64_scratch.pt |
same recipe from scratch (how much does the image warm-start buy?) — landing soon | 30k |
dit_img64_p2.pt |
the other architecture: DiT (transformer) + flow matching, patch 2 — landing soon | 30k |
Used by the Colab tutorial as the warm-start image model and the "fully-trained" reference. Generate a prompt-controlled dance:
python scripts/rollout.py --ckpt unet_t2v64.pt --prompt "A person jumps in place." --seconds 5 --n 8 --out dance.gif
Prefer standard 🤗 diffusers? See sprited/dancing-chibi-figures-ddpm-64
(unconditional DDPMPipeline) and sprited/dancing-chibi-figures-t2i-64
(text-to-image, mini-Stable-Diffusion style). Made by Sprited.