Diffusion transformers, flow matching

last edited

Diffusion transformers and flow matching are two research directions that shape current image and video models. The diffusion transformer paper (December 2022) swapped the U-Net backbone for a transformer. The flow matching paper (October 2022) defined a new training objective. Recent releases combine both ideas.

Architectural Directions

The diffusion transformer paper proposes replacing the commonly used U-Net backbone with a transformer. A CVPR 2026 paper on multi-patch transformers states that diffusion transformers are widely used in diffusion and flow-matching models. See diffusion transformer research for a longer survey.

Flow matching is a training objective for generative models built on continuous flows. FLUX.2 is a latent flow-matching model, released in November 2025. Its successor, FLUX 3, uses a flow-matching backbone.

Recent Releases

FLUX 3 launched on 23 July in early access. It is trained jointly on image, video, and audio. The FLUX 3 announcement calls it a new multimodal foundation model.

In September, the company released a robot world action model with 7B parameters and open weights. On 15 October, the open-weights press release confirmed open weights for the smaller image model. Cross-modal progress is tracked in multimodal AI research.

Open Weights and Production Uptake

Much recent work ships with weights, not only papers. The open-source AI success story of 2025-2026 covers that trend. The path from preprint to workflow node is covered in research impact on production. Weekly coverage appears on YouTube Creators & Educational Platforms.

Sources