MVDiff: Scalable and Flexible Multi-View Diffusion for 3D Object Reconstruction from Single-View

TL;DR

This paper introduces MVDiff, a novel framework that generates consistent multi-view images from a single image for 3D reconstruction, leveraging scene transformers and view-conditioned diffusion with geometric constraints.

Contribution

It presents a scalable, flexible diffusion-based method that improves 3D reconstruction quality from minimal input by incorporating geometric constraints and multi-view attention.

Findings

Outperforms baseline methods in PSNR, SSIM, and LPIPS metrics.

Generates 3D meshes from a single image input.

Enforces 3D consistency using epipolar geometry and multi-view attention.

Abstract

Generating consistent multiple views for 3D reconstruction tasks is still a challenge to existing image-to-3D diffusion models. Generally, incorporating 3D representations into diffusion model decrease the model's speed as well as generalizability and quality. This paper proposes a general framework to generate consistent multi-view images from single image or leveraging scene representation transformer and view-conditioned diffusion model. In the model, we introduce epipolar geometry constraints and multi-view attention to enforce 3D consistency. From as few as one image input, our model is able to generate 3D meshes surpassing baselines methods in evaluation metrics, including PSNR, SSIM and LPIPS.