Diffusion Transformer Research
last edited
Diffusion transformer research covers generation quality, efficiency, and uses beyond images. It is part of the key research areas feeding production tools. Recent work spans image, video, and text generation.
Quality: representation autoencoders
Representation autoencoders (RAE) pair a frozen representation encoder with a lightweight trained decoder. A diffusion transformer is then trained in that latent space. The recipe needs three additions: matching transformer width to the encoder's token dimensionality, a dimension-dependent noise-schedule shift, and decoder noise augmentation.
The arXiv paper was posted on 13 Oct 2025. It reports a large diffusion transformer with a wide DDT head. That model reaches FID 1.51 without guidance. With autoguidance it reaches FID 1.13.
The ICLR 2026 listing gives 1.18 at the smaller resolution. It gives 1.13 at the larger one. A follow-up on scaling RAEs to text-to-image appeared in Jan 2026. Its identifier is arXiv 2601.16208.
Efficiency, video, and text
ViDiT-Q was published at ICLR 2025. It quantizes diffusion transformers for image and video generation. The paper describes a 2.5x model size reduction and a 1.5x latency speedup. Both figures apply to the W4A8 setting. Such results shape research impact on production.
The diffusion-forcing transformer (DFoT) appeared at ICML 2025. It introduced history-guided video diffusion that conditions on an arbitrary number of context frames.
The LLaDA2.1 technical report was submitted in the second month of the year, on the thirteenth day. This discrete text diffusion work uses a draft-and-edit framework. The authors try to transcend the speed-quality trade-off and offer a quality mode. The model sizes are 16B and 100B, called the mini and flash variants.
Beyond media
All-atom diffusion transformers (ADiT) target 3D molecular generation. A unified latent diffusion framework jointly generates periodic materials and non-periodic molecular systems. The model matches molecule-specific and crystal-specific baselines on MP20 and QM9. It also covers GEOM-DRUGS.
For the path from paper to artist workflows, see researchers who shipped and the technique adoption timeline.
Sources
Primary papers and code for the results above are listed here.