Rethinking Cross-Layer Information Routing in Diffusion Transformers

Researchers proposed Diffusion-Adaptive Routing, a timestep-aware alternative to standard residual connections in Diffusion Transformers.

The paper finds that traditional residual addition can create redundant cross-layer information flow and hurt gradient stability.

Diffusion-Adaptive Routing learns how to aggregate sublayer outputs by timestep, improving efficiency and visual generation quality.

On ImageNet, it delivers better FID and faster convergence, and it also helps in fine-tuning and distillation setups.