Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs
LOWN cuts diffusion VLM inference by checking whether text-only predictions agree with the latest image-conditioned output, preserving quality while reducing visual processing.