PixWorld: Unified 3D Scene Generation and Reconstruction at 1000× Diffusion Speed

Researchers have introduced PixWorld, a method that unifies 3D scene generation and reconstruction by operating entirely in pixel space. The approach supports image-to-3D and text-to-3D tasks and achieves a peak speedup of roughly 1000× over diffusion-based baselines, with inference running at approximately 0.6 seconds per scene using a 4-step distilled model at 480P resolution.

PixWorld reports a WorldScore average of 71.04 and a PSNR of 26.21 dB on the RealEstate10K benchmark for 4-view reconstruction. The method covers indoor, outdoor, and stylized scenes and produces 3D Gaussian Splatting (3DGS) rendered depth outputs, positioning it as a potential accelerant for VFX and virtual production pipelines that require rapid scene asset creation.