Research map / 3D vision / 3D generation
GaussianGPT: Towards Autoregressive 3D Gaussian Scene Generation
Nicolas von Lützow, Barbara Roessle, Katharina Schmid, Matthias Niessner
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use xyz column-wise serialization instead of Z-order (ablation) | GaussianGPT · ASE+3D-FRONT | validation cross-entropy: 2.432 -> 2.421 | Improved |
| use xyz serialization instead of Hilbert curve (ablation) | GaussianGPT · ASE+3D-FRONT | validation cross-entropy: 2.439 -> 2.421 | Improved |
| use xyz serialization instead of transposed Z-order (ablation) | GaussianGPT · ASE+3D-FRONT | validation cross-entropy: 2.423 -> 2.421 | No clear effect |
| add separate vocabularies for position and feature tokens (ablation) | GaussianGPT · ASE+3D-FRONT | validation cross-entropy: 2.449 -> 2.421 | Improved |
| use 3D RoPE instead of learned positional embeddings (ablation) | GaussianGPT · ASE+3D-FRONT | validation cross-entropy: 2.461 -> 2.421 | Improved |
| use 3D RoPE instead of 1D RoPE (ablation) | GaussianGPT · ASE+3D-FRONT | validation cross-entropy: 2.446 -> 2.421 | Improved |
| reduce autoencoder downsampling from 3x to 1x (ablation) | autoencoder · 3D-FRONT reconstruction | PSNR: 21.11 -> 24.34 | Improved |
| reduce autoencoder downsampling from 3x to 2x (ablation) | autoencoder · 3D-FRONT reconstruction | PSNR: 21.11 -> 23.73 | Improved |
| increase codebook size from 4096 to 16384 (ablation) | autoencoder · 3D-FRONT reconstruction | PSNR: 21.11 -> 21.37 | Mixed |
| decrease codebook size from 4096 to 1024 (ablation) | autoencoder · 3D-FRONT reconstruction | PSNR: 21.11 -> 20.93 | Hurt |
| use standard VQ instead of LFQ (ablation) | autoencoder · 3D-FRONT reconstruction | PSNR: 21.11 -> 21.59 | Mixed |
| add resampling retries for empty columns (5 or 15 retries vs 0) (ablation) | GaussianGPT outpainting · 12m x 12m scene KID degradation | KID | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DiffRF — DiffRF: Rendering-Guided 3D Radiance Field Diffusion (CVPR 2023) · better on 1 of 1 shared result
- pi-GAN — Pi-GAN: Periodic Implicit Generative Adversarial Networks for 3D-Aware Image Synthesis (CVPR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Prometheus: 3D-Aware Latent Diffusion Models for Feed-Forward Text-to-3D Scene Generation (CVPR 2025 · similarity 0.88)
- AutoPartGen: Autoregressive 3D Part Generation and Discovery (NeurIPS 2025 · similarity 0.88)
- A 3D Scene is Worth 1K Tokens: 3D-Grounded Representation for Scene Generation at Scale (NeurIPS 2026 · similarity 0.88)
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction (CVPR 2024 · similarity 0.88)
- Atlas Gaussians Diffusion for 3D Generation (ICLR 2025 · similarity 0.88)
- UMAMI: Unifying Masked Autoregressive Models and Deterministic Rendering for View Synthesis (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
