Tessera-small: a 47.6M-param masked diffusion LM that paints pixel art in ~16 parallel steps, trained in 5.5h on one GPU

RisingSayak · x · 2026-11-19

Tessera-small treats a 16x16 sprite as a 256-token sentence: each pixel is one token from a learned 512-color palette. This masked diffusion LM starts from a fully masked canvas and unmasks all 256 pixels in parallel in 16 steps, most confident first, conditioned on CLIP text — no VAE, no latents, no autoregression.

The model has just 47.6M parameters and was trained in 5.5 hours on a single GPU over 1.4M sprites. Caleb Fahlgren built a Hugging Face Space running an int8 ONNX build in-browser via WebGPU, letting you watch the canvas fill in live — a neat proof that discrete diffusion can at least do image generation.

Original post →

More from Multimodal

Multimodal channel →