[CVPR 2026] PropFly: Learning to Propagate via On-the-Fly Supervision from Pre-trained Video Diffusion Models

Official model weights for PropFly.

PropFly is a novel training pipeline for propagation-based video editing that eliminates the need for large-scale, paired (source and edited) video datasets. Instead, it leverages on-the-fly supervision from pre-trained Video Diffusion Models (VDMs).

Model Description

Propagation-based video editing enables precise user control by propagating a single edited frame into subsequent frames while maintaining the original context. Our proposed method, PropFly, achieves this by:

1. On-the-Fly Supervision: Utilizing a frozen pre-trained VDM to synthesize structurally aligned yet semantically distinct source (low-CFG) and target (high-CFG) latent pairs on the fly.
2. Guidance-Modulated Flow Matching (GMFM): Training an adapter to learn propagation by predicting the VDM's high-CFG velocity, conditioned on the source video structure and the edited first frame style via GMFM loss.

This approach ensures temporally consistent and dynamic transformations, significantly outperforming state-of-the-art methods on various video editing tasks (evaluated on EditVerseBench and TGVE benchmarks).

Repository Structure

The model weights are stored in the PropFly-1.3B/ directory.

├── PropFly-1.3B/
│   ├── diffusion_pytorch_model.bin  # Model weights
├── .gitattributes
└── README.md

Citation

@article{seo2026propfly,
  title={PropFly: Learning to Propagate via On-the-Fly Supervision from Pre-trained Video Diffusion Models},
  author={Seo, Wonyong and Moon, Jaeho and Lee, Jaehyup and Kim, Soo Ye and Kim, Munchurl},
  journal={arXiv preprint arXiv:2602.20583},
  year={2026}
}