Making a 2D GAN 3D-Aware (original) (raw)

"What is really needed to make an existing 2D GAN 3D-aware?"

To answer this question, we modify a classical GAN, i.e., StyleGANv2, as little as possible. We find that only two modifications are absolutely necessary:

1. A multiplane image style generator branch which produces a set of alpha maps conditioned on their depth;
2. A pose-conditioned discriminator.

We refer to the generated output as a 'generative multiplane image' (GMPI) and emphasize that its renderings are not only high-quality but also guaranteed to be view-consistent, which makes GMPIs different from many prior works. Importantly, the number of alpha maps can be dynamically adjusted and can differ between training and inference, alleviating memory concerns and enabling fast training of GMPIs in less than half a day at a resolution of 10242. Our findings are consistent across three challenging and common high-resolution datasets, including FFHQ, AFHQv2 and MetFaces.

RGB and Geometry Split View

Please click to play / pause each figure; drag the separator to see the pixel-aligned geometry (during the video playing).
Click here to reset.
Desktop browser is recommended for this to work properly.

Bibtex

                    `
        @inproceedings{zhao2022gmpi,
            title = {Generative Multiplane Images: Making a 2D GAN 3D-Aware},
            author = {Xiaoming Zhao
                and Fangchang Ma
                and David Güera
                and Zhile Ren
                and Alexander G. Schwing
                and Alex Colburn},
            booktitle = {Proc. ECCV},
            year = {2022},
        }
                    `
                

Acknowledgements

Work done as part of Xiaoming Zhao's internship at Apple. Supported in part by NSF grants 1718221, 2008387, 2045586, 2106825, MRI #1725729, NIFA award 2020-67021-32799