diffcua

Differential vision encoding for computer-use agents. Instead of sending full screenshots each frame, encode only the visual diff between consecutive frames. Reduces token cost and latency for CUA inference.