# PointDiT: Pixel-Space Diffusion Model Generates Dense 3D Geometry from Single Images

_Research · published 2026-07-06_

Researchers from Google, ETH Zurich, the University of Tübingen, Microsoft, and TU Munich have introduced PointDiT, a diffusion model for monocular geometry estimation that operates directly in pixel space rather than through a lossy VAE latent bottleneck. Accepted to ICML 2026, the model takes a single input image and pure Gaussian noise and generates dense 3D point maps and depth maps in as few as one denoising step — with additional steps refining fine geometric detail.

The core insight is that both dominant paradigms for single-image 3D geometry — deterministic regression and latent-space diffusion — introduce systematic blurring. Regression heads average over ambiguous predictions, while VAE compression destroys fine structure before generation begins. PointDiT sidesteps both problems by patchifying raw XYZ point maps as tokens (analogous to RGB patches in a vision transformer) and denoising them with a plain ViT conditioned on frozen DINOv3 image features, with no separate VAE training stage required.

## Sources
- [haofeixu.github.io](https://haofeixu.github.io/pointdit/)
---
Canonical: https://genbuzz.news/posts/pointdit-pixel-space-diffusion-model-generates-dense-3d-geometry-from-single-images