How to avoid encoding the same video frame twice?

Video multimodal models waste computational resources by encoding each frame independently, even though adjacent frames are nearly identical. AdaCodec instead sends a full reference frame only when the scene changes unpredictably, otherwise transmitting compact descriptions of motion and pixel differences. Across 11 benchmarks, this adaptive approach matches Qwen3-VL-8B's performance at 1/7 the token budget and accelerates response time from 9.3 seconds to 1.6 seconds.