PRX Part 3 — Training a Text-to-Image Model in 24h!

Hugging Face · 7 months ago · 39

Researchers trained a text-to-image diffusion model in 24 hours using 32 H200 GPUs and combined multiple optimization techniques including pixel-space training, token routing, perceptual losses, and representation alignment. The total compute cost was approximately $48 at $2 per GPU-hour, down from millions of dollars for competitive models in earlier diffusion research. The resulting model produces usable images with strong prompt following at 512 and 1024 pixel resolutions, though the authors note remaining texture artifacts and anatomical inconsistencies consistent with undertraining rather than fundamental flaws.