59 upvotes
New routing cuts visual work in diffusion VLMs
What is Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs?
The paper introduces certificate-guided visual routing, which skips processing image patches that don’t affect the current denoising step. This reduces the 89-97 % visual token overhead that slows diffusion vision-language models.
Who it is for
Researchers and developers building diffusion VLMs can speed up inference without sacrificing quality, letting them run larger models or deploy on limited hardware.