SceneTok: A Compressed, Diffusable Token Space for 3D Scenes [CVPR '26]

Mohammad Asim, Christopher Wewer, Jan Eric Lenssen

Max Planck Institute for Informatics, Saarland Informatics Campus

TL;DR: A scene autoencoder that encodes a scene into a compressed, unstructured and diffusable 1D token representation.

📣 News

• 24.03.2026 - SceneGen Code, documentation and checkpoint release
• 09.03.2026 - SceneTok Code, documentation and checkpoint release
• 23.02.2026 - Initial repository release
• 21.02.2026 - Accepted to CVPR 2026 🎉!

🔍 Method Overview

(Left) SceneTok encodes view sets into a set of compressed, unstructured scene tokens by chaining a VA-VAE image compressor and a perceiver module. The tokens can be rendered from novel views with a generative decoder based on rectified flows.

(Right) SceneGen perform scene generation by generating compressed scene tokens conditioned on a single or a few images and a set of anchor poses, defining the spatial scene extent.

🔧 Installation

System Requirements

CUDA >= 11.8
Python >= 3.9
PyTorch >= 2.5

Note

We only tested on the following so far:

torch-2.5.1 + Python-3.11 + CUDA 12.1

torch 2.7.0 + Python 3.11 + CUDA 12.8

Hardware Requirements

Tested hardware for inference:

GPU	Mem
L40S	45GB
Quadro RTX 8000	48GB
RTX 4090	24GB

Test hardware for training:

GPU	Mem
A100	40GB/80GB
H100	96GB

Setting Up Environment

Clone the repository

git clone https://github.com/mohammadasim98/scenetok.git --recursive
cd scenetok

You use any environment manager, we show an example for conda

conda create -n scenetok -python=3.11
conda activate scenetok

# torch 2.5.1 + cuda 12.1
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121

Install remaining dependencies.

pip install -r requirements.txt

We use Flash Attention 2 in our project, refer to their repository for more details regarding installation.

pip install psutils ninja packaging
pip install flash-attn --no-build-isolation

Downloading Data

Datasets

Place the dataset in dataset/ folder. Refer to DATASET.md for more information into downloading and preprocessing the datasets.

Models

Download the SceneTok models to checkpoints/ folder.

Dataset	Model	rFID
RealEstate10K	va-videodc_re10k.ckpt	11.12
DL3DV	va-videodc_dl3dv.ckpt	19.12
	va-wan_dl3dv.ckpt	14.30

Download the SceneGen models to checkpoints/ folder. Note that you also need to download the corresponding va-videodc_re10k_scene.ckpt model used for the SceneGen.

Dataset	Model	gFID
RealEstate10K	scenegen_shift1_re10k.ckpt	19.99
	scenegen_shift4_re10k.ckpt	19.04
	scenegen_shift12_re10k.ckpt	18.90

Pretrained AEs

Download VA-VAE including the latent statistics to checkpoints/ folder.

Download VideoDCAE to checkpoint/ folder.

Download Wan 2.2 VAE to checkpoint/ folder.

Note

If you are using other environments than conda, then modify the environmental setup in scripts/*sh.

🏃 Quick Inference in Jupyter Notebook

We provide an experimental notebook to perform quick inference with SceneTok and SceneGen.

🤖 Training SceneTok

Run the following to initiate training:

python -m src.main +experiment=${config} \
  data_loader.train.num_workers=${num_workers} \
  mode=train \
  trainer.devices=${gpus} \
  trainer.num_nodes=${num_nodes} \
  wandb.activated=true

Important

In case of memory issues (OOM), you can reduce the overal memory consumption by setting gradient_checkpointing=true for model.denoiser and model.compressor in config/experiments/*.yaml. They are set to false by default. set environment variable DEBUG=1 for debugging purposes which disables torch.compile for all modules. Note that torch.compile is enabled by default for all modules (only used during training step). Set the parameters according to your needs. We provide the following configurations files ${config} for SceneTok. Note that you can modifiy/add your own configurations. The configuration for DL3DV finetunes pretrained weights from RealEstate10K, but you can train from scratch as well.

# RealEstate10
scenetok_va-vdc_lognorm_re10k_scratch

# DL3DV (VA-VAE + VideoDCAE)
scenetok_va-vdc_shift4_dl3dv_finetuned
scenetok_va-vdc_shift8_dl3dv_finetuned

# DL3DV (VA-VAE + Wan)
scenetok_va-wan_shift4_dl3dv_finetuned
scenetok_va-wan_shift8_dl3dv_finetuned

Using SLURM

We provide SLURM script to allow multi-node/gpu training using job arrays. In scripts/slurm_job_array_scenetok.sh we provide a template where you can define your own configuration related to GPU and job array configurations. First, set the PROJECT_ROOT as environment variable:

export PROJECT_ROOT="<your-project-directory>"

The simply run it as:

bash slurm_job_array_scenetok.sh ${config}

⚙️ Inference on SceneTok

python -m src.main +experiment=${config} mode=test hydra.job.name=test \
  dataset=${dataset} \ 
  wandb.activated=false \ 
  trainer.limit_test_batches=1 \
  data_loader.test.batch_size=1 \
  model.compressor.ckpt_path=${ckpt} \
  model.compressor.load_strict=false \
  model.denoiser.ckpt_path=${ckpt} \
  model.denoiser.load_strict=false \
  dataset.root=${data_root} \
  hydra.run.dir=${output_dir} \
  dataset/view_sampler=${view_sampler} \
  dataset.view_sampler.index_path=${index_path}

Note

model.compressor and model.denoiser uses the same checkpoint file load_strict should be set to false to allow backward compatibility for all modules.

Select view_sampler from evaluation_video for VideoDCAE-based models and evaluation_video_wan for Wan-based models.

List below are the config with checkpoints weights provided here

# RealEstate10
scenetok_va-vdc_lognorm_re10k_scratch # va-videodc_re10k.ckpt

# DL3DV (VA-VAE + VideoDCAE)
scenetok_va-vdc_shift8_dl3dv_finetuned # va-videodc_dl3dv.ckpt

# DL3DV (VA-VAE + Wan)
scenetok_va-wan_shift4_dl3dv_finetuned # va-wan_dl3dv.ckpt

index_path are view indices to use for context and targets. We provide several index files in assets/evaluation_index/

Choose dataset from re10k and dl3dv for RealEstate10K and DL3DV datasets respectively.

🤖 Training SceneGen

SceneGen generates compressed scene tokens conditioned on a single or a few images and a set of anchor poses. It requires a pretrained SceneTok model.

Run the following to initiate training:

python -m src.main_scene +experiment=${config} \
  data_loader.train.num_workers=${num_workers} \
  mode=train \
  trainer.devices=${gpus} \
  trainer.num_nodes=${num_nodes} \
  wandb.activated=true

Important

SceneGen uses src.main_scene instead of src.main as the entry point.

Set the parameters according to your needs. We provide the following configurations files ${config} for SceneGen:

# RealEstate10K
scenegen_shift1_re10k   # timestep_shift=1
scenegen_shift4_re10k   # timestep_shift=4
scenegen_shift12_re10k  # timestep_shift=12

Note

The timestep_shift parameter controls the noise schedule. Higher shift values generally lead to better generation quality. The pretrained SceneTok model va-videodc_re10k_scene.ckpt is required and automatically loaded via the config.

Using SLURM

We provide SLURM script to allow multi-node/gpu training using job arrays. In scripts/slurm_job_array_scenegen.sh we provide a template where you can define your own configuration related to GPU and job array configurations. First, set the PROJECT_ROOT as environment variable:

export PROJECT_ROOT="<your-project-directory>"

Then simply run it as:

bash scripts/slurm_job_array_scenegen.sh ${config}

⚙️ Inference on SceneGen

python -m src.main_scene +experiment=${config} mode=test hydra.job.name=test \
  dataset=re10k \
  wandb.activated=false \
  trainer.limit_test_batches=1 \
  data_loader.test.batch_size=1 \
  model.compressor.ckpt_path=${scenetok_ckpt} \
  model.compressor.load_strict=false \
  model.denoiser.ckpt_path=${scenetok_ckpt} \
  model.denoiser.load_strict=false \
  model.scene_generator.ckpt_path=${scenegen_ckpt} \
  model.scene_generator.load_strict=false \
  dataset.root=${data_root} \
  hydra.run.dir=${output_dir} \
  dataset/view_sampler=evaluation_video \
  dataset.view_sampler.index_path=${index_path}

Note

model.compressor and model.denoiser use the same SceneTok checkpoint (va-videodc_re10k_scene.ckpt), while model.scene_generator uses a separate SceneGen checkpoint. load_strict should be set to false to allow backward compatibility for all modules.

List below are the config and checkpoint combinations:

# RealEstate10K
scenegen_shift1_re10k   # scenegen_shift1_re10k.ckpt
scenegen_shift4_re10k   # scenegen_shift4_re10k.ckpt
scenegen_shift12_re10k  # scenegen_shift12_re10k.ckpt

# All SceneGen configs require:
# - scenetok_ckpt=checkpoints/va-videodc_re10k_scene.ckpt
# - scenegen_ckpt=checkpoints/scenegen_shift{1,4,12}_re10k.ckpt

We also provide a Jupyter notebook for interactive inference: SceneGen Notebook

✅ TODO

SceneTok

• RealEstate10K - VA-VAE+VideoDC
• DL3DV - VA-VAE+Wan
• DL3DV - VA-VAE+VideoDC

SceneGen

• RealEstate10K - VA-VAE+VideoDC

📣 Future Extensions

• Interactive Scene Renderer
  • Causal decoder version for SceneTok
• High-resolution SceneTok (512x512)
• Pointcloud prediction from RGB+Depth renderings
• SceneGen (DL3DV)

📃 Feedback

We will be very happy to hear feedback on how best to improve the documentation. Since we cleaned and restructured the code alot, in case of any issue or bugs with the documentation or the codebase, feel free to let us know. Refer to KNOWN_BUGS for bugs already reported.

🔖 BibTeX

If you find our work useful in your research, please cite it as:

@inproceedings{asim26scenetok,
    title = {SceneTok: A Compressed, Diffusable Token Space for 3D Scenes},
    author = {Asim, Mohammad and Wewer, Christopher and Lenssen, Jan Eric},
    booktitle = {IEEE/CVF Computer Vision and Pattern Recognition ({CVPR})},
    year = {2026},
}