# V3D: Video Diffusion Models are Effective 3D Generators

**Authors:** Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, Huaping Liu  
**arXiv:** [2403.06738](https://arxiv.org/abs/2403.06738)  
**Date:** 2024-03-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.06738v1/

## Key points

- Novel method for generating consistent multi-view images using image-to-video diffusion models
- Fine-tuning the base model on 3D datasets to extend capabilities as an effective 3D generator
- Synthesizing 360° videos of 3D objects from a single image input
- Tailored reconstruction pipeline for extracting underlying 3D asset from generated views
- Providing precise control over camera path while maintaining excellent multi-view consistency
- Extensive experiments validating effectiveness and strong generalization abilities
- Remarkable performance in generating consistent multi-views
- Code available at https://github.com/heheyas/V3D for further exploration and implementation

## Summary

In this paper, we introduce V3D, a novel method for generating consistent multi-view images using image-to-video diffusion models. Building on recent advancements in video diffusion models, we fine-tune the base model on 3D datasets to extend its capabilities as an effective 3D generator. Specifically focusing on object-centric 3D generation, we fine-tune the video diffusion model to synthesize 360° videos of 3D objects, enabling the prediction of dense views from a single image input. To extract the underlying 3D asset from these generated views, we propose a tailored reconstruction pipeline with designed initialization and texture refinement processes. This allows us to reconstruct high-quality meshes or detailed textured 3D Gaussians within just three minutes. Additionally, our framework extends to scene-level novel view synthesis, providing precise control over the camera path while maintaining excellent multi-view consistency. Extensive experiments validate the effectiveness of our approach, showcasing its remarkable performance in generating consistent multi-views and demonstrating strong generalization abilities. We believe that our method can serve as a powerful and efficient tool for high-quality 3D generation tasks and pave the way for broader applications of video diffusion models in various 3D-related tasks. Our code is readily available at https://github.com/heheyas/V3D for further exploration and implementation.