Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion

✍️ Authors

Matiur Rahman Minar¹, Seunghun Oh², Ganghyeon Jeong², Unsang Park^1,2
1Department of Computer Science and Engineering, Sogang University   ²Department of Artificial Intelligence, Sogang University

🚀 Progress

• 📝 Technical Report / Paper
• 🌐 Project Homepage
• 💻 Training & Inference Code
• 🤗 Pretrained Model: T2V-1.3B

🎯 Overview

Steady-Forcing produces long-horizon nature video rollouts from a fixed-camera view. It decouples spatial persistence from motion continuity via a structural dual-memory protocol. This enables stable backgrounds and sustained fluid motion.

TL;DR: We propose a dual-memory framework that balances stability and motion to sustain high background persistence and continuous fluid dynamics over multi-minute horizons for fixed-camera nature video generation.

📋 Table of Contents

• Requirements
• Installation
• Pretrained Checkpoints
• Inference
• Training
• Results
• Citation
• Acknowledgements

🔧 Requirements

• Nvidia GPU with at least 24 GB memory (tested on NVIDIA A100 with 80 GB VRAM)
• Linux operating system

Other hardware may work but has not been tested.

🛠️ Installation

Create a Python 3.10 environment, install dependencies, and download models:

bash setup_env.sh

📦 Pretrained Checkpoints

Download text prompts and ODE initialization checkpoint

hf download minar09/Steady-Forcing-T2V-1.3B --local-dir ./ckpt

Note: The training algorithm is data-free distillation; no video data is needed.

File Structure

After downloading, organize the checkpoints and prompts as follows:

steady-forcing/
├── prompts/
├── ckpt/
    └── steady-forcing-t2v.pt

🚀 Inference

Run inference with the provided script:

bash inference.sh

🏋️ Training

The repository can also be used for training and evaluation.

Self-Forcing training with DMD

bash train.sh

This training recipe was completed in under 67 hours on 8 A100 GPUs.

📊 Results

Quantitative and qualitative results are available in the paper. For detailed comparisons and visualizations, please refer to the arXiv preprint. For viewing generated videos, please visit the project page.

📄 Citation

If you use this codebase, please cite:

@article{minar2025steady,
  title={Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion},
  author={Minar, Matiur Rahman and Oh, Seunghun and Jeong, Ganghyeon and Park, Unsang},
  journal={arXiv preprint arXiv:2606.7661673},
  year={2026}
}

🤝 Acknowledgements

This project builds on the open-source Infinity-RoPE and Reward-Forcing implementation and acknowledges related work in long-horizon video diffusion, motion continuity, and spatial persistence. We sincerely appreciate their efforts and thank them.