Researchers introduce OpenCoF, a framework comprising the OpenCoF-17K dataset and the Wan-CoF model, to advance Chain-of-Frame (CoF) reasoning in video generation. The OpenCoF-17K dataset covers 11 task families, providing diverse temporal supervision that addresses gaps in existing general-purpose video generators.
- Wan-CoF is a fine-tuned version of the Wan2.2-I2V-A14B baseline model.
- It achieves considerable gains across four video reasoning benchmarks.
- The study explores equipping the model with visual and textual reasoning tokens to capture low-level visual cues and high-level semantic priors.
- Analysis examines how these tokens contribute across model depth, denoising steps, space, and time.
The authors conclude that stronger video reasoning requires both broad temporal supervision and explicit mechanisms for organizing intermediate reasoning state. The dataset, model, and code are open-sourced to facilitate future research.