Research map / Video / Video generation
CoF-T2I: Video Models as Pure Visual Reasoners for Text-to-Image Generation
Chengzhuo Tong, Mingkun Chang, Shenglong Zhang, Yuran Wang, Cheng Liang, Zhizheng Zhao, Bohan Zeng, Yang Shi, Ruichuan An, Yifan Dai, Ziming Zhao, Guanbin Li and 3 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune with CoF trajectories (CoF-T2I) vs base video model | Wan2.1-T2V-14B · GenEval | overall: 0.55 -> 0.86 | Improved |
| fine-tune with CoF trajectories vs base video model | Wan2.1-T2V-14B · Imagine-Bench | overall: 5.939 -> 7.468 | Improved |
| add intermediate frame supervision (full CoF vs target-only SFT) (ablation) | CoF-T2I · GenEval | overall: 0.81 -> 0.86 | Improved |
| add intermediate frame supervision (full CoF vs target-only SFT) (ablation) | CoF-T2I · Imagine-Bench | overall: 6.755 -> 7.468 | Improved |
| add target-only SFT on CoF-Evol-Instruct final frames (ablation) | Wan2.1-T2V-14B · GenEval | overall: 0.55 -> 0.81 | Improved |
| add independent frame-wise VAE encoding (ablation) | CoF-T2I · GenEval | overall: 0.83 -> 0.86 | Improved |
| add a refinement step (Frame 1 to Frame 2) (ablation) | CoF-T2I · GenEval | overall: 0.56 -> 0.79 | Improved |
| add a refinement step (Frame 2 to Frame 3) (ablation) | CoF-T2I · GenEval | overall: 0.79 -> 0.86 | Improved |
| add a refinement step (Frame 1 to Frame 3) (ablation) | CoF-T2I · Imagine-Bench | overall: 6.015 -> 7.468 | Improved |
| apply CoF-T2I recipe to 1.3B backbone (ablation) | Wan2.1-T2V-1.3B · GenEval | overall: 0.22 -> 0.79 | Improved |
| apply CoF-T2I recipe to 14B backbone | Wan2.1-T2V-14B · GenEval | overall: 0.55 -> 0.86 | Improved |
| use pure visual CoF reasoning instead of textual planning (BAGEL-Think) | CoF-T2I · GenEval | overall: 0.82 -> 0.86 | Improved |
Closest papers by meaning
- OpenCoF: Learning to Reason Through Video Generation (NeurIPS 2026 · similarity 0.92)
- T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · similarity 0.90)
- Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation (ICML 2026 · similarity 0.88)
- Breaking Dual Bottlenecks: Evolving Unified Multimodal Models into Self-Adaptive Interleaved Visual Reasoners (ICML 2026 · similarity 0.88)
- Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm (CVPR 2026 · similarity 0.88)
- GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
