Wan-CoF

Wan-CoF is the data-only model from OpenCoF: Learning to Reason Through Video Generation. It is obtained by LoRA fine-tuning Wan2.2-I2V-A14B on OpenCoF-17K, without adding a reasoning-specific architecture.

Adapter files

Both adapters are required because the base model has separate denoising experts:

high_noise_lora/model.safetensors
low_noise_lora/model.safetensors

The adapters do not include the Wan2.2-I2V-A14B base weights.

Evaluation

Headline results reported in the paper:

Benchmark	Metric	Wan2.2-I2V-A14B	Wan-CoF
MME-CoF	Overall (0–4)	1.00	1.30
VIPER	POC@1.0	3.3	7.5
RULER-Bench	Overall, I2V subset (0–100)	55.8	56.8

See the paper for the complete protocol and per-category results.

License and citation

The adapters are released under Apache-2.0 and require the separately licensed Wan2.2-I2V-A14B base model. Please cite both OpenCoF and Wan when using them.

@article{chen2026opencof,
  title   = {OpenCoF: Learning to Reason Through Video Generation},
  author  = {Chen, Xinyan and Guo, Ziyu and Zhang, Renrui and Jiang, Dongzhi and Li, Hongsheng},
  journal = {arXiv preprint arXiv:2607.08763},
  year    = {2026}
}