Wonder3D: A Look At Our Method and Consistent Multi-view Generation

Table of Links

Abstract and 1 Introduction

2. Related Works

2.1. 2D Diffusion Models for 3D Generation

2.2. 3D Generative Models and 2.3. Multi-view Diffusion Models

3. Problem Formulation

3.1. Diffusion Models

3.2. The Distribution of 3D Assets

4. Method and 4.1. Consistent Multi-view Generation

4.2. Cross-Domain Diffusion

4.3. Textured Mesh Extraction

5. Experiments

5.1. Implementation Details

5.2. Baselines

5.3. Evaluation Protocol

5.4. Single View Reconstruction

5.5. Novel View Synthesis and 5.6. Discussions

6. Conclusions and Future Works, Acknowledgements and References

4. Method

As per our problem formulation in Section 3.2, we propose a multi-view cross-domain diffusion scheme, which operates on two distinct domains to generate multi-view consistent normal maps and color images. The overview of our method is presented in Figure 2. First, our method adopts a multi-view diffusion scheme to generate multi-view normal maps and color images, and enforces the consistency across different views using multi-view attentions (see Section 4.1). Second, our proposed domain switcher allows the diffusion model to operate on more than one domain while its formulation does not require a re-training of an existing (potentially single domain) diffusion model such as Stable Diffusion [45]. Thus, we can leverage the generalizability of large foundational models, which are trained on a large corpus of data. A cross-domain attention is proposed to propagate information between the normal domain and color image domain ensuring geometric and visual coherence between the two domains (see Section 4.2). Finally, our novel geometry-aware normal fusion reconstructs the high-quality geometry and appearance from the multi-view 2D normal and color images (see Section 4.3).

4.1. Consistent Multi-view Generation

The prior 2D diffusion models [31, 45] generate each image separately, so that the resulting images are not geometrically and visually consistent across different views. To enhance consistency among different views, similar to prior works such as SyncDreamer [33] and MVDream [51], we utilize attention mechanism to facilitate information propagation across different views, implicitly encoding multi-view dependencies (as illustrated in Figure 4)

This is achieved by extending the original self-attention layers to be global-aware, allowing connections to other views within the attention layers. Keys and values from different views are connected to each other to facilitate the exchange of information. By sharing information across different views within the attention layers, the diffusion model perceives multi-view correlation and becomes capable of generating consistent multi-view color images and normal maps.