WorldForge

     

WorldForge: Unlocking Emergent 3D/4D Generation in Video Diffusion via Training-Free Guidance

Chenxi Song¹, Yanming Yang¹, Tong Zhao¹, Ruibo Li², Chi Zhang^1*

¹AGI Lab, Westlake University
²The College of Computing and Data Science, Nanyang Technological University
^*Corresponding Author

TODO

• Paper released on arXiv
• Project page available
• Code and implementation details (Coming very soon)
• Inference pipeline and usage manual

Update

• [2025.09] arXiv preprint is available.
• [2025.09] Project page is online.

Introduction

Welcome to WorldForge! WorldForge is a training-free framework that unlocks the world-modeling potential of video diffusion models, delivering controllable 3D/4D generation with unprecedented realism. Our method leverages the rich latent world priors of large-scale video diffusion models to achieve precise trajectory control and photorealistic content generation without requiring additional training.

Features

• Training-Free Framework: No additional training or fine-tuning required, preserving pretrained knowledge
• Intra-Step Recursive Refinement (IRR): Recursive refinement mechanism during inference for precise trajectory injection
• Flow-Gated Latent Fusion (FLF): Leverages optical flow similarity to decouple motion from appearance in latent space
• Dual-Path Self-Corrective Guidance (DSG): Adaptively corrects trajectory drift through guided and unguided path comparison
• 3D Scene Generation: Generate controllable 3D scenes from single view images
• 4D Video Re-cam: Dynamic trajectory-controlled re-rendering of video content
• Video Editing: Support for object removal, addition, face swapping, and subject transformation

Method Overview

WorldForge adopts a warping-and-repainting pipeline with three complementary mechanisms:

1. IRR (Intra-Step Recursive Refinement): Enables precise trajectory injection through recursive optimization
2. FLF (Flow-Gated Latent Fusion): Selectively injects trajectory guidance into motion-related channels
3. DSG (Dual-Path Self-Corrective Guidance): Maintains trajectory consistency and visual fidelity

Results

3D Scene Generation from Single View

...

4D Video Re-cam

...

Video Editing Applications

...

We showcase diverse capabilities including:

• 3D Scene Generation: Voyager experiences in artworks, AIGC content, portrait photography, city walks
• 4D Video Re-cam: Camera arc rotation, local close-ups, outpainting, viewpoint transferring, video stabilization
• Video Editing: Object removal/addition, face swapping, subject transformation, try-on applications

Installation and Usage

Prerequisites

# Installation instructions will be provided upon code release
# Requirements: Python 3.8+, PyTorch, etc.

Quick Start

# Code and detailed usage instructions coming soon
# Example usage will be provided here

Inference

# Detailed documentation will be available upon release

Comparisons

Our method demonstrates superior performance compared to existing SOTA methods:

• 3D Scene Generation: More consistent scene content under novel viewpoints with improved detail and trajectory accuracy
• 4D Video Re-cam: Realistic high-quality content re-rendering along target trajectories
• Quantitative Metrics: Superior results in realism, trajectory consistency, and visual fidelity

Citation

@misc{song2025worldforgeunlockingemergent3d4d,
  title={WorldForge: Unlocking Emergent 3D/4D Generation in Video Diffusion Model via Training-Free Guidance}, 
  author={Chenxi Song and Yanming Yang and Tong Zhao and Ruibo Li and Chi Zhang},
  year={2025},
  url={https://arxiv.org/abs/2509.15130}, 
}

Acknowledgments

We thank the research community for their valuable contributions to video diffusion models and 3D/4D generation. Special thanks to the following open-source projects that inspired and supported our work:

• Wan-2.1 - Large-scale video generation model
• SVD (Stable Video Diffusion) - Video diffusion model by Stability AI
• VGGT - 3D generation and reconstruction toolkit
• ReCamMaster - Trajectory-controlled video generation
• TrajectoryCrafter - Trajectory-based video synthesis
• NVS-Solver - Novel view synthesis solution
• ViewExtrapolator - View extrapolation for 3D scenes
• DepthCrafter - Video sequence depth estimation
• Mega-SAM - Video depth and pose estimation

Contact

For questions and discussions, please feel free to contact:

• Chenxi Song: [email protected]
• Chi Zhang: [email protected]

---

🌟 Star us on GitHub if you find WorldForge useful! 🌟