Key Info
Tessera-small is a masked diffusion language model for 16x16 pixel art that treats each sprite as a 256-token sentence, with every pixel drawn from a learned 512-color palette.
Highlights
- Starts from a fully masked canvas and unmasks all 256 pixels in parallel in about 16 steps, most confident pixels first, conditioned on CLIP text
- Uses no VAE, no latents, and no autoregression
- 47.6M parameters, trained in roughly 5.5 hours on one GPU over 1.4M sprites
- An int8 ONNX build runs in the browser on WebGPU via Caleb's Slack Moji Space, letting users watch the canvas fill in