chain-of-frame

OpenCoF introduces a reasoning video dataset and a fine-tuned video generation model that improves temporal reasoning through diverse supervision and explicit reasoning tokens, showing significant gains on four video reasoning benchmarks.