CARES: Context-Aware Resolution Selector for VLMs
Researchers have developed CARES, a Context-Aware Resolution Selector, designed to optimize image resolution for vision-language models (VLMs). This lightweight module predicts the minimum sufficient input resolution for a given image-query pair, reducing computational load and latency. By using a compact VLM to determine when a target VLM's response converges, CARES can cut compute by up to 80% while maintaining task performance across various benchmarks and VLMs. AI
IMPACT Reduces compute and latency for VLMs, potentially accelerating adoption and lowering operational costs.