Wonder3D: Evaluating the Quality of Novel View Synthesis for Different Methods

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

5.5. Novel View Synthesis

We evaluate the quality of novel view synthesis for different methods. The quantitative results are presented in Table 2, and the qualitative results can be found in Figure 3. Zero123 [31] produces visually reasonable images, but they lack multi-view consistency since it operates on each view independently. Although SyncDreamer [31] introduces a volume attention scheme to enhance the consistency of multi-view images, their model is sensitive to the elevation degrees of the input images and tends to produce unreasonable results. In contrast, our method is capable of generating images that not only exhibit semantic consistency with the input image but also maintain a high degree of consistency across multiple views in terms of both colors and geometry

5.6. Discussions

In this section, we conduct a set of studies to verify the effectiveness of our designs as well as the properties of the method.

Cross-Domain Diffusion. To validate the effectiveness of our proposed cross-domain diffusion scheme, we study the following settings: (a) cross-domain model with cross-domain attention; (b) cross-domain model without cross-domain attention; (c) sequential model rgb-to-normal: first train a multi-view color diffusion model then train a multiview normal diffusion model conditioned on the previously generated color images; (d) sequential model normal-torgb: first train a multi-view normal diffusion model then train a multi-view color diffusion model conditioned on the previously generated normal images.

As shown in (a) and (b) of Figure 7, it’s evident that the cross-domain attentions significantly enhance the consistency between color images and normals, particularly in terms of the detailed geometries of objects like the icecream and Pharaoh sculpture. From (c) and (d) of Figure 7, while the normals and color images generated by sequential models maintain some consistency, their results suffer from performance drops.

performance drops. For the sequential model rgb-to-normal, conditioning on the separately generated normal maps, the generated color images exhibit color aberrations in comparison to the input image, as shown in (c) of Figure 7. Conversely, for the sequential model normal-to-rgb, conditioning on the separately generated color images, the normal maps give unreasonable geometry, as illustrated in (d) of Figure 7. These experiments demonstrate that jointly predicting normal maps and color images through the cross-domain attention mechanism can facilitate a comprehensive perception of information from different domains. We also speculate that in the context of sequential models, the generated color images or normal maps of stage 1 may exhibit a minor domain gap to the ground truth data trained in stage 2. Therefore, compared to sequential prediction, the cross-domain approach proves to be more effective in enhancing the quality of each domain as well as the overall prediction

Multi-view Consistency. We conducted an analysis of the effectiveness of the multi-view attention mechanism, as illustrated in Figure 9. Our findings show that the multi-view attention greatly enhances the 3D consistency of the generated multi-view images, particularly for the rear views. In the absence of the multi-view attention, the color images of the rear views exhibited unrealistic predictions.

Normal Fusion. To assess the efficacy of our normal fusion algorithm, we conducted experiments using the complex lion model, which is rich in geometric details, as illustrated in Figure 8. The baseline model’s surfaces exhibited numerous holes and noises. Utilizing either the geometry-aware normal loss or the outlier-dropping loss helps mitigate the noisy surfaces. Finally, combining both strategies yields the best performance, resulting in clean surfaces while preserving detailed geometries.

Generalization. To demonstrate the generalization capability of our method, we conducted evaluations using diverse image styles, including sketches, cartoons, and images of animals, as shown in Figure 5 and Figure 10. Despite variations in lighting effects and geometric complexities among these images, our method consistently generated multi-view normal maps and color images, ultimately yielding high-quality geometries.