⚡ FlashRender ⚡

Few-Step Generative Rendering via Camera-Controlled Video MeanFlow

Byeongjun Park  ·  Byung-Hoon Kim  ·  Hyungjin Chung

---

Input Video (left)  ·  Target Retake (right)

🔥 High-quality camera-controlled video-to-video generation in just 4 sampling steps!! 🔥

📑 Table of Contents

• Setup
• Training
• Inference and Evaluation
• Acknowledgement
• Citation
• License

🔧 Setup

1. Create the environment and download checkpoints

   conda env create -f environment.yml
   conda activate flashrender
   pip install -r requirements.txt
   pip install flash_attn --no-build-isolation
   pip install --no-build-isolation git+https://github.com/mohammadasim98/met3r
   pip install --no-build-isolation -e vipe
   python download_model.py

2. Prepare the training datasets

   Download MultiCamVideo and SyncCamVideo, and place them as:

   <path/to/dataset>/
   ├── MultiCamVideo/
   │   └── train/
   │       ├── f18_aperture10/
   │       ├── f24_aperture5/
   │       ├── f35_aperture2.4/
   │       └── f50_aperture2.4/
   └── SyncCamVideo/
       └── train/
           └── f24_aperture5/
   

3. Extract features

   # Captioning is optional — ours is already provided in dataset/metadata.csv
   torchrun --nproc-per-node=8 text_captioning.py dataset_path=<path/to/dataset>
   
   torchrun --nproc-per-node=8 extract_features.py \
       dataset_path=<path/to/dataset> meta_path=dataset/metadata.csv

🚀 Training

FlashRender retakes an input video from a target camera trajectory in 4-NFE using a three-stage training recipe, run on 8 GPUs and configured through configs/base.yaml.

1. Stage 1 (train.py) — Multi-step flow-matching model (t = r)

   torchrun --nproc-per-node=8 train.py dataset_path=<path/to/dataset>

2. Stage 2 (train_meanflow.py) — Few-step MeanFlow model, resumed from Stage 1

   torchrun --nproc-per-node=8 train_meanflow.py \
       meanflow.resume_checkpoint=models/checkpoints/epoch_20.safetensors

3. Stage 3 (train_onpolicy.py) — On-policy flow-map distillation of the Stage 2 student

   torchrun --nproc-per-node=8 train_onpolicy.py \
       onpolicy.teacher_checkpoint=models/checkpoints/epoch_20.safetensors \
       onpolicy.student_checkpoint=models/checkpoints/meanflow_epoch_20.safetensors

🎬 Inference and Evaluation

1. Download checkpoints (skip if you already ran download_model.py during setup)

   Stage	Checkpoint file	Download
   1	epoch_20.safetensors	Checkpoint (Stage 1)
   2	meanflow_epoch_20.safetensors	Checkpoint (Stage 2)
   3	onpolicy_epoch_5.safetensors	Checkpoint (Stage 3)

2. Run inference

   # Stage 1: multi-step checkpoint
   python inference.py eval.ckpt_path=models/checkpoints/epoch_20.safetensors \
       eval.num_inference_steps=50 eval.inference_mode=mul \
       eval.cam_type=1 eval.video_path=example_data/bear
   
   # Stage 2: MeanFlow checkpoint
   python inference.py eval.ckpt_path=models/checkpoints/meanflow_epoch_20.safetensors \
       eval.num_inference_steps=4 eval.inference_mode=any \
       eval.cam_type=1 eval.video_path=example_data/bear
   
   # Stage 3: on-policy-distilled checkpoint
   python inference.py eval.ckpt_path=models/checkpoints/onpolicy_epoch_5.safetensors \
       eval.num_inference_steps=4 eval.inference_mode=any \
       eval.cam_type=1 eval.video_path=example_data/bear

3. Evaluate a generated video (the example below evaluates the output of the final checkpoint)

   python evaluate.py --data_path results/bear/onpolicy_epoch5_4NFE_cam1.mp4

🙏 Acknowledgement

This project builds on the following excellent works. Many thanks to the authors for releasing their code.

• ReDirector: Creating Any-Length Video Retakes with Rotary Camera Encoding
• ReCamMaster: Camera-Controlled Generative Rendering from A Single Video
• REPA: Representation Alignment for Generation
• VGGT: Visual Geometry Grounded Transformer
• Improved MeanFlow: On the Challenges of Fastforward Generative Models
• AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation
• FastGen: Fast Generation from Diffusion Models

📚 Citation

If you find this repository helpful for your project, please consider citing our work. :)

@article{park2026flashrender,
  title={FlashRender: Few-Step Generative Rendering via Camera-Controlled Video MeanFlow},
  author={Park, Byeongjun and Kim, Byung-Hoon and Chung, Hyungjin},
  journal={arXiv preprint arXiv:2609.03563},
  year={2026}
}

📄 License

This project is released under the Apache License 2.0.