Neural upsampling for higher-quality latent rendering

Apply latent-space neural upsampling to physically based latent rendering to preserve and enhance details and sharpness when generating higher-resolution outputs.

Background

The latent representations used by the method have substantially lower spatial resolution than the final RGB images. While this compression makes latent rendering efficient, it can cause aliasing when highly detailed objects are rendered.

The paper points to latent-space super-resolution as a possible way to improve the resolution of rendered latents before decoding. Because latent representations encode high-level image features in a spatially compressed form, upsampling in latent space may preserve details and sharpness more effectively than conventional RGB upsampling. The integration of such a method remains unresolved.