Abra Study Reveals Diffusion Models Need 10x More Data Than LLMs for Compute Optimal Scaling

baaadas · x · 2026-08-20

This research presents a systematic scaling law study for text-to-image diffusion models using Abra, a family of flow-matching transformers trained across three orders of magnitude of compute ($10^{19}$ to $10^{22}$ FLOPs). It demonstrates that diffusion models scale predictably like language models but require significantly more data for optimal training: compute optimality occurs at approximately 200 image tokens per parameter, ten times the Chinchilla prescription for LLMs. The study also shows diffusion models are robust to overtraining, suggesting practitioners should err on the side of more data. This predictability extends to generation quality metrics, optimal CFG settings, and representation quality.

Related event: Luma AI Publishes Scaling Laws for Text-to-Image Diffusion Models(3 posts)→

Original post →

More from Multimodal

Multimodal channel →