Search papers, labs, and topics across Lattice.
This paper introduces PixWorld, a novel model that unifies 3D scene generation and reconstruction by employing a pixel-space diffusion approach, directly supervising diffusion on rendered images. By eliminating the need for latent encoding and pretrained models, PixWorld addresses significant limitations of previous methods, including information loss and misalignment with 3D fidelity. The model not only outperforms existing latent-space generation techniques but also matches the performance of state-of-the-art reconstruction methods, showcasing the advantages of a pixel-centric strategy.
A unified pixel-space approach in PixWorld achieves superior 3D scene generation and reconstruction without the pitfalls of latent encoding.
3D reconstruction and generation are commonly tackled by separate paradigms: pixel-based regression for reconstruction, and latent diffusion for generation. Recent works attempt to unify them in latent space, but with notable drawbacks: the diffusion objective is defined on latent features rather than the underlying 3D representation, and both branches suffer from information loss introduced by latent encoding, while requiring a pretrained Variational Autoencoder (VAE) or Representation Autoencoder (RAE). In this paper, we reformulate these two tasks under a unified pixel-space diffusion paradigm and introduce PixWorld, a single model that jointly addresses 3D reconstruction and generation. By supervising diffusion directly on rendered images, PixWorld removes the above limitations and aligns optimization with 3D scene fidelity. Beyond photometric and perceptual supervision that operates at the 2D image level and lacks 3D geometric awareness, we further introduce a geometry perception loss that aligns rendered views with their ground truth in the geometry-aware feature space of a pretrained 3D foundation model, providing 3D structural supervision. PixWorld consistently outperforms prior latent-space generation methods and matches state-of-the-art reconstruction methods, demonstrating the superiority of a unified pixel-space approach.
