VGC: Learning to Correct Geometry in Generated Videos

NeurIPS 2026 | Code | Project Page

Video Geometry Corrector (VGC) is a post-hoc video-to-video model that corrects geometry distortions in generated videos (warped structures, inconsistent parallax, drifting scene layouts) without modifying the source generator.

For usage instructions, see the official code repository.

Contents

• wan22: Release version of the VGC Wan2.2-I2V-A14B MoE model (720p, no optical flow), fine-tuned from Wan2.2-I2V-A14B
  • release_no_flow_high_noise.pth: high-noise expert
  • release_no_flow_low_noise.pth: low-noise expert

Citation

@inproceedings{lyu2026vgc,
    title={Learning to Correct Geometry in Generated Videos},
    author={Lyu, Weijie and Shen, Tiancheng and Li, Xiangtai and Wang, Yujing and Wang, Xiaoyuan and Yuan, Haobo and Zhao, Yizhou and Yang, Ming-Hsuan},
    booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
    year={2026}
}