sprite32: open-weights 32x32 text-to-sprite diffusion model trained from scratch

DaneMarguglio · reddit · 2026-11-18

A developer trained sprite32 from scratch — a 143M-parameter UNet for 32x32 transparent-background sprites, trained on 660k sprites/icons/emoji in 20 hours on a 5090. It works directly on pixels (no VAE) and outputs RGBA. Two versions: 1.0 (9GB VRAM, 0.9s/sprite, tunable guidance) and 1.0 fast (5x faster at 0.16s, under 3GB, CPU-friendly, 2.3GB download). The whole project was vibe-coded by Claude over 5-6 sprints. Weights Apache-2.0 on Hugging Face, demo at sprite32.com.

Original post →

More from Multimodal

Multimodal channel →