FOCUS Reduces VLM Video Inference via Temporal Gating
November 3, 2026
FOCUS is a training-free framework that optimizes vision-language model inference by using temporal gating to trigger captures only during semantic scene changes. It uses a spatial focus map to calibrate the VLM's value cache, reducing the frequency of full frame processing.
HOW THIS AFFECTS YOU
●
builderYou can reduce the high inference costs of real-time video processing without retraining your models.
●
designerThis enables more responsive and efficient multimodal interactions in video-based applications.