Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NeRF-VAE: A Geometry Aware 3D Scene Generative Model

Adam R. Kosiorek, Heiko Strathmann, Daniel Zoran, Pol Moreno, Rosalia Schneider, Soňa Mokrá, Danilo J. Rezende

TL;DR

NeRF-VAE addresses 3D scene generation by embedding a geometry-aware NeRF-based decoder within a variational framework. It introduces a view-invariant latent z and a conditional scene function to capture shared scene priors, enabling rapid amortized inference for novel scenes and view-consistent rendering via differentiable volume rendering. The approach demonstrates improved generalization to out-of-distribution cameras and the ability to model uncertainty from limited context, with an attention-based conditioning mechanism enhancing performance. The work highlights the benefits and limitations of combining explicit geometry with learned priors and points toward future work on dynamic scenes and richer latent representations.

Abstract

We propose NeRF-VAE, a 3D scene generative model that incorporates geometric structure via NeRF and differentiable volume rendering. In contrast to NeRF, our model takes into account shared structure across scenes, and is able to infer the structure of a novel scene -- without the need to re-train -- using amortized inference. NeRF-VAE's explicit 3D rendering process further contrasts previous generative models with convolution-based rendering which lacks geometric structure. Our model is a VAE that learns a distribution over radiance fields by conditioning them on a latent scene representation. We show that, once trained, NeRF-VAE is able to infer and render geometrically-consistent scenes from previously unseen 3D environments using very few input images. We further demonstrate that NeRF-VAE generalizes well to out-of-distribution cameras, while convolutional models do not. Finally, we introduce and study an attention-based conditioning mechanism of NeRF-VAE's decoder, which improves model performance.

NeRF-VAE: A Geometry Aware 3D Scene Generative Model

TL;DR

NeRF-VAE addresses 3D scene generation by embedding a geometry-aware NeRF-based decoder within a variational framework. It introduces a view-invariant latent z and a conditional scene function to capture shared scene priors, enabling rapid amortized inference for novel scenes and view-consistent rendering via differentiable volume rendering. The approach demonstrates improved generalization to out-of-distribution cameras and the ability to model uncertainty from limited context, with an attention-based conditioning mechanism enhancing performance. The work highlights the benefits and limitations of combining explicit geometry with learned priors and points toward future work on dynamic scenes and richer latent representations.

Abstract

We propose NeRF-VAE, a 3D scene generative model that incorporates geometric structure via NeRF and differentiable volume rendering. In contrast to NeRF, our model takes into account shared structure across scenes, and is able to infer the structure of a novel scene -- without the need to re-train -- using amortized inference. NeRF-VAE's explicit 3D rendering process further contrasts previous generative models with convolution-based rendering which lacks geometric structure. Our model is a VAE that learns a distribution over radiance fields by conditioning them on a latent scene representation. We show that, once trained, NeRF-VAE is able to infer and render geometrically-consistent scenes from previously unseen 3D environments using very few input images. We further demonstrate that NeRF-VAE generalizes well to out-of-distribution cameras, while convolutional models do not. Finally, we introduce and study an attention-based conditioning mechanism of NeRF-VAE's decoder, which improves model performance.

Paper Structure

This paper contains 42 sections, 8 equations, 18 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Neural Radiance Fields (nerf)
  3. Nerf-vae
  4. Amortized Inference for
  5. Iterative Amortized Inference
  6. Conditioning the Scene Function
  7. Related Work
  8. Experiments
  9. Datasets
  10. GQN
  11. CLEVR
  12. Jaytracer
  13. Implementation details
  14. Comparison with Nerf
  15. Comparison with a Convolution-Based Generative Model
  16. ...and 27 more sections

Figures (18)

  • Figure 1: During training, we embed common scene properties (across the dataset) in the parameters $\phi, \theta$ of OURS.
  • Figure 2: Once the model is trained, we can infer parameters of a scene in constant time, even from a single view.
  • Figure 3: We can sample novel scenes from OURS generative model and render them from various viewpoints.
  • Figure 5: Inference in .
  • Figure 6: The generative model of .
  • ...and 13 more figures