LOWN: Certificate-Guided Visual Routing for Efficient Diffusion VLMs
October 26, 2026
LOWN reduces decoding computation in diffusion vision-language models by deciding at each step whether to process visual tokens. Tests on LLaDA-V show that removing the image changes token unmasking in only 22.8% of steps on average, enabling significant efficiency gains without losing information.
HOW THIS AFFECTS YOU
●
builderYou can reduce inference costs and latency for diffusion-based multimodal models.
●
researcherThis provides a new method for optimizing visual token processing in generative workflows.