# PixGS: Single-Stage Pixel-Space Diffusion Generates 3D Gaussian Splats in One Second

_Research · published 2026-07-10_

Researchers Cao Duy and Phong Nguyen-Ha have introduced PixGS, a single-stage diffusion pipeline for generating 3D Gaussian Splats (3DGS) directly from text or images. Unlike prevailing approaches that adapt latent diffusion models through multi-stage cascade pipelines, PixGS operates in pixel space, sidestepping the lossy compression and accumulated decoding artifacts that constrain existing methods. The model denoises 3D Gaussian attributes at each timestep and applies supervision over surface normals, depth, and high-frequency structural detail.

Accepted at ECCV 2026, the paper reports that PixGS outperforms current state-of-the-art multi-stage methods while achieving inference in approximately one second on a single A100 GPU. The approach is notable for maintaining access to large-scale 2D generative priors while eliminating the scalability bottlenecks inherent in cascaded architectures — a combination that could accelerate practical 3D asset generation workflows in VFX and virtual production.

## Sources
- [arxiv.org](https://arxiv.org/abs/2607.01803)
---
Canonical: https://genbuzz.news/posts/pixgs-single-stage-pixel-space-diffusion-generates-3d-gaussian-splats-in-one-second