Why this model exists: it is the answer sheet for learning image generation on sprited/dancing-chibi-figures. The dataset is meant to be the first dataset you ever train a diffusion model on — and this checkpoint is what you get when you follow the standard 🤗 diffusers recipe on it, with nothing clever added: the official unconditional-training tutorial architecture (UNet2DModel + DDPMScheduler, squaredcos, EMA), 30k steps, ~71M params. If your run looks like the grid below, you did it right. It doubles as proof that the dataset trains cleanly with textbook code, and as the unconditional floor that the conditional models (text-to-image, video) are measured against.

One small twist worth knowing: it diffuses 4-channel premultiplied RGBA directly (no VAE) — the model learns the transparent background as part of the image.

Training script (deliberately mirrors the official diffusers unconditional-training tutorial): train_diffusers/train_ddpm.py. Want text control? See sprited/dancing-chibi-figures-t2i-64. Made by Sprited.