📐 PointDiT: Accurate 3D Geometry from a Single Photo
The PointDiT model has been developed to estimate 3D geometry from a single monocular image. Unlike latent space methods (VAE), PointDiT applies diffusion (Flow Matching) directly in pixel space (raw point maps), which eliminates the blurring of fine structures.
🌍 The transition to pixel-space diffusion solves the "bottleneck" problem in geometry reconstruction, allowing for the creation of high-precision depth and normal maps without losing small objects.
👤 This is a critical step for robotics and augmented reality, as systems will be able to understand complex 3D scenes containing thin and transparent objects from a standard photo.
Source 1: https://haofeixu.github.io/pointdit/