🌍 $I^2$-World: Intra-Inter Tokenization for Efficient Dynamic 4D Scene Forecasting

September 5, 2025 · View on GitHub

[2025-06] $I^2$ -World is accepted to ICCV 2025.

🛠️Environment

Install Pytorch 1.13 + CUDA 11.6

conda create --name ii-world python=3.8
conda activate ii-world
pip install torch==1.13.0+cu116 torchvision==0.14.0+cu116 torchaudio==0.13.0 --extra-index-url https://download.pytorch.org/whl/cu116

Install mmdet3d (v1.0.0rc4) related packages and build this project

pip install mmcv-full==1.7.0 -f https://download.openmmlab.com/mmcv/dist/cu117/torch1.13/index.html
pip install mmdet==2.28.2
pip install mmsegmentation==0.30.0
pip install mmengine
pip install -v -e .

Install other dependencies

pip install -r requirements.txt

🤗 Model Zoo

We utilize 8 RTX4090 GPUs to train our model.

Method	Dataset	Task	Rec.mIoU (%)	Rec.IoU (%)	Weights
II-Tokenizer	Occ3D-nus	Rec	81.1	68.1	Google-drive
	STCOcc-Res	Rec	24.8	32.2	-
	Occ3D-Waymo	Rec	76.3	74.6	-
II-World	Occ3D-nus	Fore	38.4	49.2	Google-drive
	STCOcc-Res	Fore	18.9	28.8	-
	Occ3D-Waymo	Fore	43.7	60.9	-

📦 Prepare Dataset

Download nuScenes from nuScenes
Download Occ3D-nus from Occ3D-nus
(Optional) Download Occ3D-Waymo from Occ3D-Waymo and unzip it to the data/waymo folder. We only use the validation of Occ3D-Waymo in our experiments.
(Optional) Download STCOcc-Res from STCOcc-Res and unzip it to the data/nuscenes folder.
Download the generated info file from Google Drive and unzip it to the data/nuscenes folder. These *pkl files can be generated by running the tools/create_data.py
(Optional) Download the visualization car model Google Drive
Organize your folder structure as below:

├── project
├── visualizer/
│   ├── 3d_model.obj/ (optional)
├── ckpts/
│   ├── ii_scene_tokenizer_4f.pth
│   ├── ii_generate_world.pth
├── data/
│   ├── nuscenes/
│   │   ├── samples/ 
│   │   ├── v1.0-trainval/
│   │   ├── gts/ (Occ3D-nus)
│   │   ├── stc-results/ (prediction from STCOcc) (optional)
│   │   ├── world-nuscenes_infos_train.pkl
│   │   ├── world-nuscenes_infos_val.pkl
│   ├── waymo(optional)/
│   │   ├── validation/ 
│   │   ├── cam_infos_vali.pkl/ 
│   │   ├── waymo_infos_val.pkl/

Before training or evaluating II-World, you should first evaluate the II-Tokenizer to generate the prediction tokens. By default, the II-Tokenizer will save the prediction tokens to data/nuscenes/save_dir/token_4f folder.

Also, make sure you utilize the number of GPU can be divided by 150 when evaluating the II-Tokenizer.

You can change the test_data_config in the tokenizer config for different datasets.

When generate the training set prediction tokens, you can set the ann_file in test_data_config to world-nuscenes_infos_train.pkl

Train II-World with 8GPUs:

bash tools/dist_train.sh configs/world_model/ii_generate_world.py 8

Evaluate II-World with 6GPUs:

bash tools/dist_test.sh configs/world_model/ii_generate_world.py TO/CKPTS

🎥 Visualization

We provide a simple visualization to visualize the high-level control (utilize different cmd) of the world generation.

python tools/generate.py configs/world_model/ii_generate_world.py ckpts/ii_generate_world.pth \
--generate_path generate_output --generate_scene_name scene-0564 --generate_frame 12 --task_mode high-level-control

Also, you can visualize the generated world with the fine-grained control (utilize different transformation matrix)

python tools/generate.py configs/world_model/ii_generate_world.py ckpts/ii_generate_world.pth \
--generate_path generate_output --generate_scene_name scene-0270 --generate_frame 12 --task_mode generate

If you want to visualize the 3D occupancy map, you can set --save_npz in the script above, and the generated 3D occupancy npz will be saved in the generate_output folder.

Utilize the following command to visualize the generated 3D occupancy map:

python tools/vis_occ_3d.py --vis-single-data \PATH/TO/GENERATED/3D_OCCUPANCY.npz --vis-path demo_output

More visualization options can be found in the tools/vis_occ_3d.py file.

Acknowledgement

Thanks to the following excellent projects:

🌍 $I^2$-World: Intra-Inter Tokenization for Efficient Dynamic 4D Scene Forecasting

🌍 $I^2$ -World: Intra-Inter Tokenization for Efficient Dynamic 4D Scene Forecasting

arXiv

🚀 News

🛠️Environment

🤗 Model Zoo

📦 Prepare Dataset

🎇 Training and Evaluation

🎥 Visualization

Acknowledgement