cd /news/artificial-intelligence/gaussiangpt-generating-3d-scenes-aut… · home topics artificial-intelligence article
[ARTICLE · art-114933] src=github.com ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

GaussianGPT: Generating 3D scenes autoregressively with Gaussian Splatting

Researchers from the Technical University of Munich introduced GaussianGPT, a transformer-based model that generates 3D scenes autoregressively via next-token prediction on Gaussian splatting tokens, accepted as an oral presentation at ECCV 2026. The model compresses 3D Gaussians into discrete tokens using a sparse 3D convolutional VQ-VAE and generates scenes step-by-step, supporting completion, outpainting, and controllable sampling. The full training and inference code, along with pre-trained checkpoints, were released on GitHub.

read12 min views1 publishedAug 29, 2026
GaussianGPT: Generating 3D scenes autoregressively with Gaussian Splatting
Image: Michielbdejong (auto-discovered)

Nicolas von Lützow · Barbara Rössle · Katharina Schmid · Matthias Nießner

Project Page · arXiv · Paper · Video

Most recent advances in 3D generative modeling rely on diffusion or flow-matching formulations. We instead explore a fully autoregressive alternative and introduce GaussianGPT, a transformer-based model that directly generates 3D Gaussians via next-token prediction, thus facilitating full 3D scene generation. We first compress Gaussian primitives into a discrete latent grid using a sparse 3D convolutional autoencoder with vector quantization. The resulting tokens are serialized and modeled using a causal transformer with 3D rotary positional embedding, enabling sequential generation of spatial structure and appearance. Unlike diffusion-based methods that refine scenes holistically, our formulation constructs scenes step-by-step, naturally supporting completion, outpainting, controllable sampling via temperature, and flexible generation horizons. This formulation leverages the compositional inductive biases and scalability of autoregressive modeling while operating on explicit representations compatible with modern neural rendering pipelines, positioning autoregressive transformers as a complementary paradigm for controllable and context-aware 3D generation.

2026-08-05- GaussianGPT selected for an oral presentation!** 2026-07-12**- Object-level (PhotoShape)checkpointsreleased.** 2026-07-02**- Camera-ready paper now available onarXiv.** 2026-07-01**- Pre-trained scene-level VQ-VAE and GPTcheckpointsreleased.** 2026-06-19**- Training and inference code released.** 2026-06-18**- GaussianGPT accepted to ECCV 2026!

This repository contains the full training and inference code for GaussianGPT. The pipeline is two trained models connected by a tokenization step:

VQ-VAE(train_ae.py

) — a sparse 3D CNN with vector quantization that compresses per-voxel Gaussians into discrete tokens, supervised by agsplat

re-rendering loss.Tokenization(tokenize_dataset.py

) — runs the trained encoder over the dataset and writes per-scene token streams to disk.GPT(train_gpt.py

) — an autoregressive transformer with 3D rotary embeddings that models the token streams via next-token prediction.Inference— sample from the GPT and decode through the frozen VQ-VAE to generate, complete, or tile Gaussian scenes, then render them.

Gaussians --VQ-VAE--> tokens --GPT--> sampled tokens --decode--> Gaussians --render-->

Everything is configured with Hydra; override any field on the CLI as key=value

.

If you find GaussianGPT useful, please consider citing:

@inproceedings{vonluetzow2026gaussiangpt,
  title     = {GaussianGPT: Towards Autoregressive 3D Gaussian Scene Generation},
  author    = {von L{\"u}tzow, Nicolas and R{\"o}{\ss}le, Barbara and Schmid, Katharina and Nie{\ss}ner, Matthias},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
}

The environment is built around CUDA 12.9 and PyTorch 2.8. Several dependencies are compiled from source, so make sure a GPU is visible during installation for correct CUDA support.

Set the target architectures before any from-source build. Refer to the NVIDIA GPU feature list for your hardware.

export TORCH_CUDA_ARCH_LIST="8.6;8.0"   # e.g. Ampere (3090, A6000, A100)
conda create -n gaussiangpt python=3.10 nvidia::cuda-toolkit=12.9 conda-forge::glm ninja
conda activate gaussiangpt

ln -s "$CONDA_PREFIX/lib" "$CONDA_PREFIX/lib64"
conda env config vars set CUDA_HOME="$CONDA_PREFIX"
export CUDA_HOME="$CONDA_PREFIX"

pip install torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu129
pip install --no-build-isolation -r requirements.txt   # compiles several extensions; slow

requirements.txt

installs two CUDA extensions from pinned git commits that are compiled from source during this step (hence the GPU-visible requirement and the long build time): gsplat (the re-rendering loss / renderer) and

(rotation/quaternion ops in

pytorch3d

utils/transforms.py

). If the install fails, it is almost always one of these two — build them individually to see the full nvcc error, and make sure TORCH_CUDA_ARCH_LIST

and CUDA_HOME

are set as above.

git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention
git checkout v2.8.2
MAX_JOBS=4 python setup.py install

Troubleshooting: MAX_JOBS=1 #

Each nvcc job needs several GB of RAM. Start MAX_JOBS

low and raise it only if you have the RAM to spare; lower it (down to 1) if the build OOMs / gets "Killed".

Optional: FLASH_ATTN_CUDA_ARCHS=80 #

flash-attn compiles its own arch set (sm_80;90;100;120) and ignores TORCH_CUDA_ARCH_LIST

. Its Ampere kernels are sm_80, which also run on sm_86 (3090/A6000) via same-major forward-compat, so restricting to sm_80 covers all the target GPUs above and cuts both compile time and per-job RAM substantially.

The upstream MinkowskiEngine does not build against CUDA 12 without source patches. The alpsaur fork bundles the CUDA 12 fixes, so no manual patching is needed:

git clone https://github.com/alpsaur/MinkowskiEngine.git --branch cuda12-compat
cd MinkowskiEngine
conda install -c conda-forge "blas=*=openblas" openblas openblas-devel
python setup.py install --blas_include_dirs="${CONDA_PREFIX}/include" --blas=openblas

Compiler note: build with GCC <= 13 #

MinkowskiEngine does not compile with GCC >= 14 against PyTorch 2.8's bundled pybind11 — you'll hit ambiguous template instantiation

errors on the enum_

registrations. Use GCC <= 13 (GCC 11 is known good). The openblas-devel

install above can pull a newer GCC from conda-forge into the env, so after running it check ${CXX} --version

; if it reports 14+, pin an older toolchain and reactivate so conda repoints CC

/CXX

before building:

conda install -c conda-forge gcc_linux-64=11 gxx_linux-64=11
conda deactivate && conda activate gaussiangpt   # reactivate to refresh CC/CXX

If conda balks at re-solving cuda-toolkit

, pin it explicitly and freeze the rest: conda install -c nvidia -c conda-forge cuda-toolkit=12.9.2 gcc_linux-64=11 gxx_linux-64=11 --freeze-installed

.

Alternative: official repo with manual patches #

Clone upstream and apply the CUDA 12 source patches yourself, following issue #543:

git clone https://github.com/NVIDIA/MinkowskiEngine.git
cd MinkowskiEngine
conda install -c conda-forge "blas=*=openblas" openblas openblas-devel
python setup.py install --blas_include_dirs="${CONDA_PREFIX}/include" --blas=openblas

Optional: per-voxel dedup (torch_scatter #

)

Not required for the default pipeline. Only needed if you enable model.make_unique=true

(collapse each voxel to its highest-opacity Gaussian), which is off in the shipped configs. The import is lazy, so the dependency is only touched when that option is turned on. In our tests it gave a negligible quality gain for the added compute.

pip install --no-build-isolation torch-scatter -f "https://data.pyg.org/whl/torch-2.8.0+cu129.html"

We are grateful to the authors of the following datasets, whose data made this work possible. Please refer to the respective sources for licensing and download instructions.

Dataset Source data Gaussians
PhotoShape

original dataSceneSplat-49koriginal data(no longer available)## 3D-FRONT availability

The original source data is no longer available, but more recent re-releases (e.g. this one) should work similarly.

Data format #

PhotoShape— standard Inria-style 3DGS.ply

files (y-up).3D-FRONT— PyTorch dicts; easiest to follow the data- code directly. The dicts hold the default 3DGS attributes, but split position intoanchor

(3D anchor positions) andoffset

.f_dc

/f_rest

hold the SH coefficients; all other attributes are storedpre-activation(logits).

Voxel-GS preprocessing #

Voxel-GS is the simplified Scaffold-GS from L3DG — no MLP, one Gaussian per voxel, no hierarchy.

PhotoShape— L3DG Voxel-GS as-is (paper Secs. 3.2.1, 3.4).** 3D-FRONT**— same, with: point cloud from back-projected depth maps; no scene normalization; 2.5 cm voxels; anchor densification off; 60k iterations; SH degree capped at 1; scales bounded by2 * voxel_size * sigmoid()

.

Training is a two-stage pipeline (VQ-VAE, then GPT) with a tokenization step in between. Checkpoints and logs land under experiment.log_dir

(logs/

by default), organized by experiment.name

.

The dataset configs reference scene-name lists under data_splits/

. Two standalone helpers under scripts/

build them in two steps — scan once to produce a per-scene stats CSV, then filter that CSV into train/val splits (re-run the cheap second step with different thresholds without re-scanning):

python scripts/dataset_quick_stats.py \
    --data-root <gaussians_root> --transforms-root <transforms_root> \
    --output logs/stats.csv

python scripts/dataset_split_from_quick_stats.py \
    --input logs/stats.csv \
    --train-split data_splits/train.txt --val-split data_splits/val.txt \
    --min-images 100 --max-points 5000000 \
    --min-extent-x 3.2 --max-extent-x 25 \
    --min-extent-y 3.2 --max-extent-y 25 --max-extent-z 4

All filters default to off; pass --help

on either script for the full set.

Trains the sparse-CNN autoencoder with vector quantization that compresses per-voxel Gaussians into discrete tokens. Reconstruction is supervised by a re-rendering loss (gsplat

).

python train_ae.py \
    data=vfront_houses \
    experiment.name=my_vqvae
  • Top-level config: conf/vqvae.yaml

(data=vfront_houses

,model=vqvae_cnn

,training=vqvae

). Swap the dataset withdata=photoshape

,data=ase

,data=spp_v2

, etc. - Loss weights live in conf/training/vqvae.yaml

; the defaults target VFront/ASE, and the file notes the PhotoShape overrides. - PhotoShape additionally needs model=vqvae_photoshape

(finer voxels, view-dependent color). The training overrides are listed at the top of that config. max_epochs

is a deliberate overestimate — stop the run by picking a checkpoint rather than waiting for it to finish.- Resume with experiment.checkpoint_path=<ckpt> experiment.continue_mode=resume

(orweights_only

to load only the weights and reset the optimizer/scheduler).

Runs the trained VQ-VAE encoder over the dataset and writes per-scene token streams to disk. The output directory becomes data.data_path

for the GPT stage, and the VQ-VAE checkpoint becomes data.vqvae_path

.

python tokenize_dataset.py \
    data=vfront_houses \
    experiment.checkpoint_path=<vqvae.ckpt> \
    training.tokenization.output_dir=<tokens_dir>
  • Shares the conf/vqvae.yaml

config; onlyexperiment.checkpoint_path

andtraining.tokenization.output_dir

are required (both are asserted at startup). training.tokenization.sort_by

controls latent ordering;training.tokenization.generate_augmented_samples=true

writes 8 variants per scene (4 z-rotations x mirrored/not).

Trains the autoregressive transformer prior over the VQ tokens. The frozen VQ-VAE is loaded from data.vqvae_path

and used only for decoding during the inline evaluation.

python train_gpt.py \
    data=tokenized_vfront \
    data.data_path=<tokens_dir> \
    data.vqvae_path=<vqvae.ckpt> \
    experiment.name=my_gpt
  • Top-level config: conf/gpt.yaml

(data=tokenized_vfront

,model=gpt

,training=gpt

). Usedata=tokenized_vfront_ase

for the combined VFront+ASE model.data_path

andvqvae_path

are required (???

) in the tokenized data configs and must be supplied. - For the object-level model use data=tokenized_photoshape model=gpt_photoshape

(GPT-2-small on a 32^3 grid). - Transformer size is set inline via the gpt_size

block inconf/model/gpt.yaml

(n_embd

,n_layer

,n_head

,n_kv_head

). The default (n_embd=1024

,n_layer=24

) matches GPT-2-medium. - Multi-GPU is auto-detected: the run uses ddp

when more than one GPU is visible. - Evaluation runs inline viaEvaluateCallback

, cadenced bytraining.output.render_frequency

.training.output.eval_data_config

selects which rawconf/data/*.yaml

is composed for the rendering. experiment.continue_run=true

auto-finds the latest checkpoint from a prior run with the samelog_dir

/name

.

Pre-trained VQ-VAE and GPT checkpoints are hosted at kaldir.vc.cit.tum.de/gaussiangpt

(sizes and SHA256 checksums in the served README). Each GPT must be paired with the VQ-VAE listed alongside it.

wget https://kaldir.vc.cit.tum.de/gaussiangpt/vqvae_vfront.ckpt
wget https://kaldir.vc.cit.tum.de/gaussiangpt/gpt_vfront.ckpt

wget https://kaldir.vc.cit.tum.de/gaussiangpt/vqvae_both.ckpt
wget https://kaldir.vc.cit.tum.de/gaussiangpt/gpt_both.ckpt

wget https://kaldir.vc.cit.tum.de/gaussiangpt/vqvae_photoshape.ckpt
wget https://kaldir.vc.cit.tum.de/gaussiangpt/gpt_photoshape.ckpt

Pass them to any inference entry point as checkpoint=<gpt.ckpt> vqvae_checkpoint=<vqvae.ckpt>

(see Inference).

All inference entry points take the trained GPT via checkpoint=<gpt.ckpt>

and need a VQ-VAE checkpoint to decode sampled tokens. The VQ-VAE is resolved as vqvae_checkpoint=<vqvae.ckpt>

(explicit override) → model.vqvae.checkpoint_path

data.vqvae_path

from a composed data=<cfg>

. The token streams (data.data_path

) are only read when conditioning on real scenes (completion); unconditional sampling does not touch them.

Samples one chunk per scene, decodes through the VQ-VAE, and renders camera trajectories. This module also provides the inline helpers imported by train_gpt.py

. For object-level models (PhotoShape) a chunk is a whole object, so this is the full object-generation pipeline; the completion and multi-chunk entry points below are scene-level.

python generate_chunks.py \
    checkpoint=<gpt.ckpt> \
    vqvae_checkpoint=<vqvae.ckpt> \
    num_samples=4 temperature=1.0

Key options (see conf/generate_chunks.yaml

, the config generate_chunks.py

loads): num_samples

, batch_size

, temperature

/top_k

/top_p

, seed

, store_samples

, render_gifs

(set render_gifs=false

to skip GIF rendering, e.g. to only dump samples via store_samples=true

). Outputs go to output_dir

(outputs/eval

by default).

Prompt-conditioned completion (sequence-prefix sanity check) #

generate_chunks.py

can also condition on a real scene by enabling completion

: it keeps the first prompt_fraction

of the token sequence and continues it, producing one completion per scene. This is a quick sanity check on the prior, used primarily during training. For completion cut by spatial extent, use complete_chunks.py

below.

python generate_chunks.py \
    checkpoint=<gpt.ckpt> \
    data=tokenized_vfront \
    data.data_path=<tokens_dir> data.vqvae_path=<vqvae.ckpt> \
    completion.enabled=true completion.prompt_fraction=0.5

The prompts are read from the tokenized data=<cfg>

(completion.split

selects train/val); prompt_fraction

is a continuous float (default 0.5).

Conditions on part of a scene and samples the rest — the autoregressive analogue of inpainting/outpainting. The prompt is cut by spatial extent (e.g. keep half the room along x) rather than by sequence length, and several completions are sampled per scene.

python complete_chunks.py \
    checkpoint=<gpt.ckpt> \
    data=tokenized_vfront \
    data.data_path=<tokens_dir> data.vqvae_path=<vqvae.ckpt> \
    prompt_mode=spatial_half_x

Key options (see conf/complete_chunks.yaml

): split

, prompt_mode

, num_completions

, num_samples

, the sampling params, and render_gifs

/store_tokens

. Supports sharding via shard_id

/num_shards

. Outputs go to outputs/complete_chunks

.

generate_scene.py

autoregressively tiles many chunks into a large scene, then decode_scene.py

turns the saved token sidecars into renderable Gaussian payloads. generate_scene.py

shards over the tile grid via the GAUSS_SHARD_ID

/ GAUSS_NUM_SHARDS

env vars, so it runs cleanly as a SLURM array — each task only reads/writes its own rank_XXXX

shard.

GAUSS_SHARD_ID=0 GAUSS_NUM_SHARDS=1 python generate_scene.py \
    checkpoint=<gpt.ckpt> \
    vqvae_checkpoint=<vqvae.ckpt> \
    num_scenes=4 output_dir=<scene_out>

python decode_scene.py \
    --output-dir <scene_out> \
    --vqvae-checkpoint <vqvae.ckpt>

generate_scene.py

can decode and render top-down inline (decode_outputs

, render_topdown

, both on by default); see conf/generate_scene.yaml

for the tiling (scene_cols_x/y

), the bootstrap/outpainting sampling params, and the empty-column handling. decode_scene.py

infers the GPT checkpoint from the run manifest, so only --output-dir

and --vqvae-checkpoint

are required.

Standalone renderers operate on decoded scene .pt

payloads (keys coords

, sh0

, opacities

, scales

, quats

, plus optional sh

with higher-order SH coefficients, rendered when present):

python render_topdown.py --input <scene.pt> --quantile 75 --resolution 1024

python render_orbit_batch.py --input <scene_or_dir> --output-dir <render_out>

python render_topdown.py --input <sample.pt> --up y --quantile 100
python render_orbit_batch.py --input <samples_dir> --output-dir <render_out> --up y --move-back 1.75

Payloads keep their dataset's native up axis: 3D-FRONT/ASE are z-up (the default), PhotoShape is y-up (pass --up y

to rotate before rendering). render_topdown.py

drops points above --quantile

(to see through ceilings; use --quantile 100

to keep whole objects). render_orbit_batch.py

accepts a single .pt

or a directory (--max-files

caps how many it processes) and writes per-frame images plus a GIF under --output-dir

. Its default --move-back

orbits from inside a room; use ~1.5-3 to orbit around an object.

To inspect a payload in an external viewer, scripts/convert_pt_to_ply.py

converts a Gaussian .pt

/.pth

payload to an INRIA-style .ply

.

This work would not have been possible without the following open-source projects, and we thank their authors and contributors.

This project is released under the MIT License. See LICENSE for details.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @technical university of munich 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/gaussiangpt-generati…] indexed:0 read:12min 2026-08-29 ·