Reward Forcing:
Efficient Streaming Video Generation with
Rewarded Distribution Matching Distillation

Yunhong Lu^1,2, Yanhong Zeng², Haobo Li^2,4, Hao Ouyang², Qiuyu Wang², Ka Leong Cheng²,
Jiapeng Zhu², Hengyuan Cao¹, Zhipeng Zhang⁵, Xing Zhu², Yujun Shen², Min Zhang^1,3

¹ZJU, ²Ant Group, ³SIAS-ZJU, ⁴HUST, ⁵SJTU

🚀 Progress

• 📝 Technical Report / Paper
• 🌐 Project Homepage
• 💻 Training & Inference Code
• 🤗 Pretrained Model: T2V-1.3B
• 🔜 Pretrained Model: T2V-14B (In progress)

🎯 Overview

TL;DR: We propose Reward Forcing to distill a bidirectional video diffusion model into a 4-step autoregressive student model that enables real-time (23.1 FPS) streaming video generation. Instead of using vanilla distribution matching distillation (DMD), Reward Forcing adopts a novel rewarded distribution matching distillation (Re-DMD) that prioritizes matching towards high-reward regions, leading to enhanced object motion dynamics and immersive scene navigation dynamics in generated videos.

📋 Table of Contents

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

🔧 Requirements

• GPU: NVIDIA GPU with at least 24GB memory for inference, 80GB memory for training.
• RAM: 64GB or more recommended.
• Linux operating system.

🛠️ Installation

Step 1: Clone the repository

git clone https://github.com/JaydenLyh/Reward-Forcing.git
cd Reward-Forcing

Step 2: Create conda environment

conda create -n reward_forcing python=3.10
conda activate reward_forcing

Step 3: Install dependencies

pip install -r requirements.txt
pip install flash-attn --no-build-isolation

Step 4: Install the package

pip install -e .

📦 Pretrained Checkpoints

Download Links

Model	Download
VideoReward	Hugging Face
Wan2.1-T2V-1.3B	Hugging Face
Wan2.1-T2V-14B	Hugging Face
ODE Initialization	Hugging Face
Reward Forcing	Hugging Face

File Structure

After downloading, organize the checkpoints as follows:

checkpoints/
├── Videoreward/
│   ├── checkpoint-11352/
│   └── model_config.json
├── Wan2.1-T2V-1.3B/
├── Wan2.1-T2V-14B/
├── Reward-Forcing-T2V-1.3B/
└── ode_init.pt

Quick Download Script

pip install "huggingface_hub[cli]"

# Download all checkpoints
bash download_checkpoints.sh

🚀 Inference

Quick Start

# 5-seconds video inference
python inference.py \
    --num_output_frames 21 \
    --config_path configs/reward_forcing.yaml \
    --checkpoint_path checkpoints/Reward-Forcing-T2V-1.3B/rewardforcing.pt \
    --output_folder videos/rewardforcing-5s \
    --data_path prompts/MovieGenVideoBench_extended.txt \
    --use_ema

# 30-seconds video inference
python inference.py \
    --num_output_frames 120 \
    --config_path configs/reward_forcing.yaml \
    --checkpoint_path checkpoints/Reward-Forcing-T2V-1.3B/rewardforcing.pt \
    --output_folder videos/rewardforcing-30s \
    --data_path prompts/MovieGenVideoBench_extended.txt \
    --use_ema

🏋️ Training

Multi-GPU Training

# bash train.sh
torchrun --nnodes=1 --nproc_per_node=8 --rdzv_id=5235 --rdzv_backend=c10d  \
    --rdzv_endpoint=$MASTER_PORT train.py  --config_path configs/reward_forcing.yaml \
    --logdir logs/reward_forcing \
    --disable-wandb

Multi-Node Training

torchrun --nnodes=$NODE_SIZE --nproc_per_node=8 --node-rank=$NODE_RANK --rdzv_id=5235 --rdzv_backend=c10d  \
    --rdzv_endpoint=$MASTER_IP:$MASTER_PORT train.py  --config_path configs/reward_forcing.yaml \
    --logdir logs/reward_forcing \
    --disable-wandb

Configuration Files

Training configurations are in configs/:

• default_config.yaml: Default configuration
• reward_forcing.yaml: Reward Forcing configuration

📊 Results

Quantitative Results

Performance on VBench

Method	Total Score	Quality Score	Semantic Score	Params	FPS
SkyReels-V2	82.67	84.70	74.53	1.3B	0.49
MAGI-1	79.18	82.04	67.74	4.5B	0.19
NOVA	80.12	80.39	79.05	0.6B	0.88
Pyramid Flow	81.72	84.74	69.62	2B	6.7
CausVid	82.88	83.93	78.69	1.3B	17.0
Self Forcing	83.80	84.59	80.64	1.3B	17.0
LongLive	83.22	83.68	81.37	1.3B	20.7
Ours	84.13	84.84	81.32	1.3B	23.1

Qualitative Results

Visualizations can be found in our Project Page.

📄 Citation

If you find this work useful, please consider citing:

@article{lu2025reward,
  title={Reward Forcing: Efficient Streaming Video Generation with Rewarded Distribution Matching Distillation},
  author={Lu, Yunhong and Zeng, Yanhong and Li, Haobo and Ouyang, Hao and Wang, Qiuyu and Cheng, Ka Leong and Zhu, Jiapeng and Cao, Hengyuan and Zhang, Zhipeng and Zhu, Xing and others},
  journal={arXiv preprint arXiv:2512.04678},
  year={2025}
}

🙏 Acknowledgements

This project is built upon several excellent works: CausVid, Self Forcing, Infinite Forcing, Wan2.1, VideoAlign

We thank the authors for their great work and open-source contribution.

📧 Contact

For questions and discussions, please:

• Open an issue on GitHub Issues
• Contact us at: yunhonglu@zju.edu.cn