Post by Johannes Künzel

The paper of my colleague Florian Barthel just got accepted at NeurIPS Recent 3D GANs based on Gaussian splatting suffered from a 3D inconsistency, causing identity shifts when rendering different camera views. CGS-GAN removes the view-conditioning and overcomes the inexorably destabilized GAN-training by introducing a multi-view regularization and a new generator architecture. Additionally, this enables output resolutions of up to 2048x2048! Check out the website: https://lnkd.in/dmbnsTKv Co-authors: Wieland Morgenstern Paul Hinzer Anna Hilsmann Peter Eisert Fraunhofer Heinrich Hertz Institute HHI Humboldt-Universität zu Berlin