dancing-chibi-figures-t2i-64

Why this model exists: it is the finished version of the model the Colab tutorial has you build — a miniature Stable Diffusion: the same standard parts (UNet2DConditionModel + a frozen CLIP ViT-B/32 text encoder + classifier-free guidance), minus the VAE, diffusing 64px RGBA pixels directly. It is the first model in the dancing-chibi-figures family that reads the actual sentence (via cross-attention) rather than a class label. 90M UNet params, 30k steps, EMA.

One column per prompt: jump / dance / cartwheel / kneel / walk / raise hands / kick / sit.

from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("sprited/dancing-chibi-figures-t2i-64",
                                         custom_pipeline="sprited/dancing-chibi-figures-t2i-64",
                                         trust_remote_code=True).to("cuda")
imgs = pipe(["a person jumps in place"] * 4, num_inference_steps=50, guidance_scale=3.0).images  # RGBA PIL

guidance_scale=0 ignores the text entirely; ~3 is a good default.

Honest limitations

• Trained on 143 unique motion prompts. A novel sentence is pulled toward the most similar trained one — this is closer to "knows its 143 moves by name" than to open-vocabulary understanding. (More prompt diversity is the v0.2 plan.)
• Prompts describe a whole motion, not one pose: "does a cartwheel" legitimately samples any split-second of a cartwheel, so the same prompt yields different poses.
• Structural check (silhouette oracle v0, n=256 vs 512 real): one_piece 100% / stray 0.0% / fg 0.116 — identical to the real floor (100% / 0.0% / 0.119). Prompt adherence is not yet scored systematically.

Training script: train_diffusers/train_t2i.py. Siblings: ddpm-64 (unconditional diffusers) · baselines (pure-PyTorch, incl. video). Made by Sprited.