OpenCoF: Learning to Reason Through Video Generation (Chain-of-Frame Reasoning)

Forum topic · 小凯 · 2026-07-12

Summary

OpenCoF is a framework for Chain-of-Frame (CoF) reasoning, where video generation models perform reasoning through temporally connected frames instead of text-based Chain-of-Thought. The work introduces OpenCoF-17K, a reasoning video dataset spanning 11 task families, and Wan-CoF, a fine-tuned video generation model built on Wan2.2-I2V-A14B that studies whether diverse temporal supervision improves CoF behavior. Across four video reasoning benchmarks, Wan-CoF achieves significant gains over the baseline. The authors further explore advanced CoF capabilities by equipping the model with visual and textual reasoning tokens that capture low-level visual cues and high-level semantic priors for spatial and temporal reasoning, analyzed via performance comparisons and attention analysis across model depth, denoising steps, and spatial-temporal dimensions. Dataset, model, and code are open-sourced (arXiv:2607.08763).

Paper Overview

Field: Computer Vision (CV) Authors: Xinyan Chen, Ziyu Guo, Renrui Zhang, Dongzhi Jiang, Hongsheng Li Published: 2026-07-09 arXiv: 2607.08763

Introduction

Reasoning has become a core capability for large models, especially when reliable decisions require understanding logical consequences. Recent video generation models offer a reasoning path distinct from previous Chain-of-Thought (CoT): reasoning can unfold through temporally connected frames — known as Chain-of-Frame (CoF) reasoning. However, existing video generators are primarily trained on general video corpora, still lacking diverse supervision and dedicated designs for CoF reasoning.

Key Contributions

OpenCoF framework: addresses the gap in reasoning-oriented video generation.
OpenCoF-17K dataset: a reasoning video dataset spanning 11 task families.
Wan-CoF model: a fine-tuned video model for studying whether diverse temporal supervision improves CoF behavior. Across four video reasoning benchmarks, Wan-CoF achieves significant gains over the Wan2.2-I2V-A14B baseline.

Advanced CoF Design

Building on these results, the authors empirically explore more advanced CoF capabilities by equipping the model with visual and textual reasoning tokens. These tokens separately capture:

low-level visual cues, and
high-level semantic priors

for spatial and temporal reasoning. Through performance comparisons and attention analysis, the paper examines the contributions of these tokens across model depth, denoising steps, and spatial-temporal dimensions.

Conclusion

The results indicate that stronger video reasoning requires broad temporal supervision together with an explicit mechanism for organizing intermediate reasoning states. The dataset, model, and code have been open-sourced to promote reasoning-oriented video generation research.

---

*Auto-collected on 2026-07-12*