PRX Part 3: 24h Text-to-Image Model Training — H200 GPUs

AI Impact Summary

PRX Part 3 demonstrates a 24-hour text-to-image model training run utilizing a combination of advanced techniques, including pixel-space training with x-prediction, TREAD token routing, and representation alignment via REPA and DINOv3. This experiment highlights the rapid advancements in diffusion model training, achieving a usable model in a single day with a modest compute budget of $1500 on three H200 GPUs. The resulting model, trained on synthetic datasets, shows promising results with consistent aesthetic and prompt following, though further refinement through increased compute and data diversity is anticipated.