Diffusion vision language models skip image processing when not needed
Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs
Computer Vision and Pattern RecognitionArtificial IntelligenceMachine Learning
Summary
Diffusion vision-language models usually look at the whole input image all the time, which takes a lot of computer power. The authors found that these models often don’t actually need the image every step of their thinking process. They made a new method called LOWN that predicts when the image matters and only uses it then, saving lots of work. This method keeps almost all the original quality but runs much faster by not repeatedly processing the same image.
What this means in practice
- •For mobile app developers: Speed up visual question answering in apps by reducing image processing steps without losing accuracy using adaptive token unmasking.
- •For cloud ai service providers: Lower the computational cost for running diffusion vision-language models on servers by skipping image use when predictions agree.
Authors
Donghyeon Cho, Joowon Kim, Seungho Shin, Eunho Yang
Abstract
Diffusion vision-language models (DVLMs) process their entire input at every denoising step. Since visual tokens make up 89-97% of this input on the benchmarks we study, most of the decoding computation is spent repeatedly processing the same image. Visual token pruning methods partially relieve this burden by removing part of the visual tokens, but they lose the removed information and still process the remaining image at every step. To measure how often the image is actually needed, we compare LLaDA-V's predictions with and without the image at every decoding step and find that removing the image changes a token being unmasked at only 22.8% of steps on average and at as few as 7.9%. We therefore propose LOWN, a decoding method that decides at every step whether to process the image and how many tokens to unmask. LOWN predicts the masked tokens without the image and processes the image only when this prediction disagrees with the last image-conditioned prediction. Agreement certifies the earlier prediction for reuse, and the number of certified tokens sets how many are unmasked in that step. We further observe that the need for the image differs across benchmarks, so LOWN's adaptive decision is needed rather than a fixed schedule. Without training or a tuned threshold, LOWN retains 99.6% of LLaDA-V's score on ten benchmarks at 35.3% of the decoding FLOPs on average. LOWN also generalizes across LLaDA-based DVLMs, composes orthogonally with visual token pruning, and runs under key-value caching for further efficiency gains. Code and project page are available at https://hyeon9698.github.io/LOWN/