Faster training for image models using diffusion transformers

Researchers found a way to make diffusion transformer models train more efficiently by automatically selecting which internal processing steps to use, rather than manually testing every option. This reduces training time while keeping the same quality output, making a powerful image-generation technique practical for more labs and companies.

If diffusion models become cheaper to train, more organizations can build and customize them, which shifts who can participate in AI development from well-funded labs toward broader research communities.