Self-Refining Video Sampling

Official Pytorch Implementation of Self-Refining Video Sampling

[Preprint 2026]-Self-Refining Video Sampling
Sangwon Jang^*, Taekyung Ki^*, Jaehyeong Jo^*, Saining Xie, Jaehong Yoon^†, Sung Ju hwang^†
(* indicates equal contribution, † indicates equal advising)

Generated by Wan2.2-A14B T2V

A parkour athlete runs up a vertical wall, grabs the ledge, and muscles up to stand on the roof in one fluid motion.	A sprinter explodes out of the starting blocks, body at a 45-degree angle, transitioning into an upright running posture.	A gymnast on a pommel horse swings their legs in wide circles (flares), supporting their entire weight on alternating hands.

Overview

This repo is organized into two main branches:

self-refine-video
├─ (current)        : Wan Series (Based on Diffusers)
└─ cosmos2.5-predict: Cosmos 2.5 predict code

Dependencies

2026.02.01: 🚨 There is a Wan model loading error with transformers==5.0.0. Please use transformers==4.57.3 until this issue is fixed.

Follow the model card for environment details:

• https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B-Diffusers

Minimum dependencies (install in your environment):

• diffusers
• transformers
• torch
• pip install opencv-python imageio imageio-ffmpeg

Usage

Edit prompts and hyperparameters in inference_pnp.py (T2V) or inference_i2v_pnp.py (I2V), then run the script.

Bibtex

@article{jang2026selfrefining,
      title={Self-Refining Video Sampling}, 
      author={Sangwon Jang and Taekyung Ki and Jaehyeong Jo and Saining Xie and Jaehong Yoon and Sung Ju Hwang},
      year={2026},
      journal={arXiv preprint arXiv:2601.18577},
}

🙏 Acknowledgements

This work builds upon:

• Diffusers - Wan pipeline
• Cosmos2.5 - Github