Abra: Scaling Diffusion Image Training
This work presents a systematic scaling law study for text-to-image diffusion models using Abra, a controlled family of flow-matching transformers trained across three orders of magnitude worth of compute, demonstrating that diffusion models scale just as predictably as language models but require far more data to train optimally.
Kyle R. Chickering, Wei-An Lin, Swayam Bhanded et al.
· 0 citations