PixGS

A minimal PyTorch implementation of PixGS: Pixel-Space Diffusion for Direct 3D Gaussian Splat Generation, which is a single-stage pixel-space diffusion model that directly generates 3D Gaussian Splats from text or image conditioning. This repository is a compact, self-contained reimplementation intended for experimentation and understanding of the core ideas from the paper.

Features

• Flow Matching objective
• Multi-view DiT backbone
• Pixel-aligned Gaussian attribute representation
• Plücker ray conditioning
• Differentiable Gaussian splat renderer
  • Pure PyTorch renderer
  • Optional gsplat backend
• Multi-stage training schedule
• RGB, depth, normal, LPIPS, and LoG supervision
• Toy synthetic dataset for end-to-end testing

Included

• End-to-end training pipeline
• Sampling
• PLY export
• Smoke test
• Procedural toy dataset (ToyBlobs)

Running

Smoke test:

python pixgs.py --smoke

Train on the toy dataset:

python pixgs.py

Use the CUDA gsplat renderer:

python pixgs.py --renderer gsplat

Outputs are written to out_pixgs_mini/.

Reference

PixGS: Pixel-Space Diffusion for Direct 3D Gaussian Splat Generation

Duy Cao, Phong Nguyen-Ha

https://arxiv.org/abs/2607.01803v2