Diffusion Transformers (DiTs) have become a leading architecture for high-quality image and video generation, but their computational cost remains a major obstacle. The new arXiv preprint "MoDDiT: Simple Token Routing in Diffusion Transformers" identifies model throughput as the key bottleneck, noting that real-time video generation, interactive world models, and on-device image synthesis are "almost a reality" yet held back by speed.
The paper introduces MoDDiT, a method that uses simple token routing to improve efficiency. While the abstract does not provide implementation details, the title and framing suggest a lightweight approach to reducing the computational burden of DiTs by selectively routing tokens through the network. The authors position this as a direct response to the throughput problem that currently limits practical deployment.
If successful, MoDDiT could help close the gap between the high sample quality of DiTs and the speed required for interactive and real-time applications. The work is at the preprint stage, so no experimental results or comparisons are available in the abstract, but the direction is clear: making Diffusion Transformers fast enough for the next generation of generative tools.