EG3D: Efficient Geometry-aware 3D Generative Adversarial Networks

Abstract

Unsupervised generation of high-quality multi-view-consistent images and 3D shapes using only collections of single-view 2D photographs has been a long-standing challenge. Existing 3D GANs are either compute-intensive or make approximations that are not 3D-consistent; the former limits quality and resolution of the generated images and the latter adversely affects multi-view consistency and shape quality. In this work, we improve the computational efficiency and image quality of 3D GANs without overly relying on these approximations. For this purpose, we introduce an expressive hybrid explicit-implicit network architecture that, together with other design choices, synthesizes not only high-resolution multi-view-consistent images in real time but also produces high-quality 3D geometry. By decoupling feature generation and neural rendering, our framework is able to leverage state-of-the-art 2D CNN generators, such as StyleGAN2, and inherit their efficiency and expressiveness. We demonstrate state-of-the-art 3D-aware synthesis with FFHQ and AFHQ Cats, among other experiments.

Videos

Qualitative results

The following videos demonstrate scene synthesis with our method, which produces both high-quality, multi-view-consistent renderings and detailed geometry.

Video 1: Color video renderings of scenes produced by our method, created by moving the camera along a path while fixing the latent code that controls the scene.

Video 2: Renderings of surfaces generated by our method, which are obtained from the density field of our 3D representation with isosurface extraction.

Interpolation

Our method inherits the qualities of the StyleGAN2 backbone, including a well-behaved latent space. The following video shows interpolation between selected points in FFHQ.

Video 3: Interpolations between latent vectors with FFHQ.

Tri-plane Representation

Training a GAN with neural rendering is expensive, so we use a hybrid explicit-implicit 3D representation in order to make neural rendering as efficient as possible. Our representation combines an explicit backbone, which produces features aligned on three orthogonal planes, with a small implicit decoder. Compared to a typical multilayer perceptron representation, our 3D representation is more than seven times faster and uses less than one sixteenth as much memory. In using StyleGAN2 as the backbone of our representation, we inherit the qualities of the backbone, including a well-behaved latent space.

Super-resolution & Dual Discrimination

We perform volumetric rendering at a moderate resolution ( 128 x 128 ) and leverage 2D image-space convolutions (super-resolution) to increase the final output resolution and image quality. Crucially, we ensure consistency between the final output image and the neural rendering (dual discrimination), which prevents view-inconsistent convolutional layers from introducing undesirable artifacts.

Video 4: Two video sequences comparing the super-resolution output (left half of each scene) to the neural volume rendering (right half of each scene).

Training natively lets the convolution super-resolution layers introduce view-inconsistent effects, such as subtle expression warping at the corners of the mouth. Adding dual discrimination (right) ensures the final renderings are consistent with the raw neural volume renderings, which helps suppress these artifacts.

Video 5: Side-by-side comparison of models trained without dual discrimination (left) and with dual discrimination (right).

Inversion

We apply the prior over 3D faces learned by our method to single-image 3D reconstruction. We use Pivotal Tuning Inversion to invert test images and recover 3D shapes and novel views.

Video 6: Single image 3D reconstruction using Pivotal Tuning Inversion. Input image (left) and reconstruction (right).

Realtime Demomonstration

An efficient architecture enables scene synthesis and rendering at at real-time framerates, opening the door for many exciting interactive applications.

Video 7: A demonstration of our method synthesizing and rendering scenes in real-time.

Split Color/Geometry Demo

Click to play / pause each figure; drag the separator to see the pixel-aligned geometry. Click here to reset.

Additional Visual Results

We include additional results in the following video.

Video 8: Additional sequences of interpolation within FFHQ and static scenes of FFHQ and AFHQv2.

Citation

@inproceedings{Chan2021,
            author = {Eric R. Chan and Connor Z. Lin and Matthew A. Chan and Koki Nagano and Boxiao Pan and Shalini De Mello and Orazio Gallo and Leonidas Guibas and Jonathan Tremblay and Sameh Khamis and Tero Karras and Gordon Wetzstein},
            title = {Efficient Geometry-aware {3D} Generative Adversarial Networks},
            booktitle = {arXiv},
            year = {2021}
          }

License

Images, text and video files on this site are made freely available for non-commercial use under the Creative Commons CC BY-NC 4.0 license . Feel free to use any of the material in your own work, as long as you give us appropriate credit by mentioning the title and author list of our paper.

Acknowledgments

We thank David Luebke, Jan Kautz, Jaewoo Seo, Jonathan Granskog, Simon Yuen, Alex Evans, Stan Birchfield, Alexander Bergman, and Joy Hsu for reviewing early drafts and for the helpful suggestions and feedback. We thank Alex Chan, Giap Nguyen, and Trevor Chan for help with figures and diagrams. Koki Nagano and Eric Chan were partially supported by DARPA’s Semantic Forensics (SemaFor) contract (HR0011-20-3-0005). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government. Distribution Statement "A" (Approved for Public Release, Distribution Unlimited). We base this website off of the StyleGAN3 website template.