Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion (2406.03184v1)

Abstract: Existing single image-to-3D creation methods typically involve a two-stage process, first generating multi-view images, and then using these images for 3D reconstruction. However, training these two stages separately leads to significant data bias in the inference phase, thus affecting the quality of reconstructed results. We introduce a unified 3D generation framework, named Ouroboros3D, which integrates diffusion-based multi-view image generation and 3D reconstruction into a recursive diffusion process. In our framework, these two modules are jointly trained through a self-conditioning mechanism, allowing them to adapt to each other's characteristics for robust inference. During the multi-view denoising process, the multi-view diffusion model uses the 3D-aware maps rendered by the reconstruction module at the previous timestep as additional conditions. The recursive diffusion framework with 3D-aware feedback unites the entire process and improves geometric consistency.Experiments show that our framework outperforms separation of these two stages and existing methods that combine them at the inference phase. Project page: https://costwen.github.io/Ouroboros3D/

• The paper's main contribution is unifying multi-view synthesis and 3D reconstruction in a recursive diffusion framework to mitigate data bias.
• It leverages a 3D-aware feedback mechanism with canonical coordinate maps to ensure synthesized views align with accurate geometric structures.
• Experimental results show higher PSNR and SSIM scores with lower LPIPS, outperforming state-of-the-art methods in image-to-3D generation.

Summary of "Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion"

"Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion" presents a unified approach to single image-to-3D object creation by integrating multi-view image generation and 3D reconstruction into a recursive diffusion process. This method leverages a self-conditioning mechanism, enabling joint training of the two stages for robust and geometrically consistent 3D generation.

The traditional methods for image-to-3D creation typically operate in two discrete stages: multi-view image synthesis and subsequent 3D reconstruction. These stages, when trained separately, introduce significant data bias during inference, thus impairing the quality of the reconstructed results. Ouroboros3D circumvents this issue by embedding both stages into a recursive diffusion framework, thereby optimizing the generation process through continuous feedback between the multi-view images and the evolving 3D model.

Methodology

The Ouroboros3D framework implements a video diffusion model for multi-view image generation and a feed-forward model for 3D reconstruction. The video diffusion model generates multi-view images by leveraging camera control for precise positional encoding at the pixel level. The 3D reconstruction model used in this paper is the Large Multi-View Gaussian Model (LGM), which operates on the concept of 3D Gaussian Splatting, allowing for efficient and high-quality 3D representation.

Key to the Ouroboros3D framework is the 3D-aware feedback mechanism. This mechanism involves incorporating rendered color images and geometric maps from the reconstruction module back into the multi-view denoising process. By utilizing canonical coordinates maps (CCM) as conditional inputs, the model ensures that the multi-view images align well with the geometric structure, thereby enhancing the consistency and details across generated views.

Experimental Results

Experiments conducted on the GSO dataset demonstrate that Ouroboros3D surpasses traditional two-stage methods and other state-of-the-art techniques that attempt to integrate multi-view generation and 3D reconstruction at the inference phase. Quantitative metrics such as PSNR, SSIM, and LPIPS indicate substantial improvements in both multi-view image quality and the geometric fidelity of the 3D reconstructions.

When evaluated against existing methods like SyncDreamer, SV3D, and VideoMV in the image-to-multi-view task, Ouroboros3D showed superior performance, achieving higher PSNR and SSIM scores and lower LPIPS values. In the image-to-3D task, it outperformed models like TripoSR, LGM, and InstantMesh, indicating that joint training with 3D feedback significantly enhances both image fidelity and geometric accuracy.

Implications and Future Directions

The Ouroboros3D framework offers several theoretical and practical implications for the field of computer vision and 3D reconstruction. By unifying the multi-view generation and 3D reconstruction stages, the approach mitigates data bias issues and leverages 3D-aware feedback for enhanced geometric consistency. This method has the potential to significantly improve applications in virtual reality, gaming, and digital content creation, where high-quality 3D models are essential.

Future research could explore the extension of Ouroboros3D to handle more complex scenarios such as dynamic scenes and real-time applications. Additionally, integrating other 3D representations, such as mesh-based models, could broaden the applicability of the framework in various industries.

In conclusion, Ouroboros3D demonstrates a novel and effective approach to image-to-3D generation by integrating and jointly training multi-view and 3D reconstruction stages. The recursive diffusion process with 3D-aware feedback profoundly improves the consistency and quality of the generated 3D models, marking a significant step forward in the field.