What if cameras only captured what matters?

High-resolution cameras generate too much data to process efficiently. This work trains a policy that dynamically decides where a dual-stream sensor should allocate its limited pixel budget during image capture—maintaining low-res context while high-res details go to regions the system predicts are relevant for the task. On multiple perception benchmarks and a real 200-megapixel camera, the approach matches baseline performance under strict bandwidth limits, closing the loop between sensing and task requirements.