ICCV High-Score Paper | ReCamMaster Gains International Attention, Offering a New Perspective on Hollywood Blockbusters}

The first author, Bai Jianhong, a PhD student at Zhejiang University, specializes in video generation and is currently seeking full-time industry positions.

As a video enthusiast, have you ever been limited by equipment when trying to achieve dynamic camera effects? For example, wanting to shoot a landscape from above without a drone, or being dissatisfied with shaky footage affecting the quality? As an AI video creator, are you content with the generated content but dissatisfied with the camera movements?

To address these issues, the research team from KeLun WanWei proposed ReCamMaster, a video generation model that allows re-aiming existing videos along new camera trajectories. Users can upload any video and specify new camera paths to re-animate the footage. The project also released a high-quality MultiCamVideo-Dataset with synchronized multi-camera videos, along with open-source code for training and testing.

In addition, ReCamMaster has strong application potential in 4D scene reconstruction, video stabilization, autonomous driving, and embodied intelligence.

• Paper Title: ReCamMaster: Camera-Controlled Generative Rendering from A Single Video
• Project Homepage: https://jianhongbai.github.io/ReCamMaster
• Code Repository: https://github.com/KwaiVGI/ReCamMaster
• Paper Link: https://arxiv.org/abs/2503.11647

1. ReCamMaster Capabilities Showcase

a) Video Re-aiming

b) 4D Scene Reconstruction

c) Video Denoising

d) Data Generation for Autonomous Driving and Embodied Intelligence

ReCamMaster’s generated videos can maintain scene consistency and dynamics, showing good generalization across different scenarios. More examples are available on the project homepage: https://jianhongbai.github.io/ReCamMaster/

2. ReCamMaster Innovations

The main innovations include:

• Introducing a new simple and effective video conditioning paradigm, significantly outperforming previous methods.
• Releasing a high-quality multi-camera synchronized video dataset with practical applications in controllable video generation and 4D reconstruction.
• Achieving near-production-level performance in re-aiming from a single video, demonstrating the huge potential of video generation models for such tasks.

3. ReCamMaster Algorithm Explanation

The core innovation of ReCamMaster is a new video conditioning paradigm, which involves concatenating the condition video and target video along the temporal dimension after patchify, showing significant performance improvements over channel-wise concatenation.

4. MultiCamVideo Dataset

The MultiCamVideo dataset is rendered using Unreal Engine 5, containing 13,600 scenes shot from 10 cameras along different trajectories, totaling 136,000 videos with 112,000 unique camera paths. It features 66 characters, 93 actions, and 37 high-quality 3D environments. An example is shown here:

5. ReCamMaster Experimental Results

Comparisons with state-of-the-art methods show that ReCamMaster significantly outperforms baselines in various metrics. Details can be found here:

6. Summary

This paper introduces ReCamMaster, a video re-aiming model that re-animates input videos along new camera trajectories. Its main innovation is a simple yet effective video conditioning paradigm, outperforming baseline methods. Additionally, the team released the MultiCamVideo-Dataset for controllable video generation and 4D reconstruction research. More details can be found in the original paper.