PixWorld unifies 3D generation and reconstruction

Pixel-space diffusion avoids VAE bottlenecking, which is a major step forward for high-fidelity 3D representation, though direct pixel supervision comes at a higher computational cost. Direct pixel-space optimization eliminates VAE encoding artifacts, resulting in significantly crisper generated outputs and cleaner geometry. Furthermore, incorporating a 3D foundation model's geometry perception loss solves the multi-view inconsistency and lack of depth awareness typical of purely 2D-supervised models. However, adapting this framework to larger-scale environments will require substantial optimization due to the resource-intensive nature of rendering-based loss calculations.