Wonder3D: Learn More About Diffusion Models

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

3. Problem Formulation

3.1. Diffusion Models

Diffusion models [22, 52] are first proposed to gradually recover images from a specifically designed degradation process, where a forward Markov chain and a Reverse Markov chain are adopted.

Given a sample z0 drawn from the data distribution p(z), the forward process of denoising diffusion models yields a sequence of noised data {zt | t ∈ (0, T)} with zt = αtz0 + σtϵ, where ϵ is random noise drawn from distribution N (0, 1), and αt, σt are fixed sequence of the noise schedule. The forward process will be iteratively applied to the target image until the image becomes complete Gaussian noise at the end.

On the contrary, the reverse chain then is employed to iteratively denoise the corrupted image, i.e., recovering zt−1 from zt by predicting the added random noise ϵ. The readers can refer to [22, 52] for more details about image diffusion models.