Vid2World: Crafting Video Diffusion Models to Interactive World Models (ICLR 2026)

 

This is the official code base for the paper Vid2World: Crafting Video Diffusion Models to Interactive World Models.

Give it a star 🌟 if you find our work useful!

🔥 News & Updates

• 🚩 2026-01: Vid2World has been accepted by ICLR 2026, congrats!

• 🚩 2025-12: We release all model checkpoints on 🤗 Huggingface.

• 🚩 2025-12: We release code for training, inference and evaluation.

📋 TL;DR

We repurpose internet-scale pretrained video diffusion models into interactive world models:

• ⚙️ Converts non-causal video diffusion backbones into autoregressive, temporally causal architectures with frame-level action conditioning.
• 🦸 Enables high-fidelity, action-conditioned video simulation and scalable world model learning across robot manipulation, 3D game simulation, and open-world navigation.

🚀 QuickStart

⚙️ Environment Setup

First create your conda environment:

conda create -n v2w python=3.8 -y
conda activate v2w

Then, install dependencies:

pip install -r requirements.txt

For training and evaluation:

• Download the base video model (DynamiCrafter, 320 $\times$ 512), and save it into checkpoints/dynamicrafter_512_v1/model.ckpt.
• Download the pretrained i3d model and save it into checkpoints/i3d/i3d_torchscript.pt.

At this point, your checkpoints folder should look like this:

checkpoints
├── dynamicrafter_512_v1
│   └── model.ckpt
└── i3d
    └── i3d_torchscript.pt

🤗 Models

At the moment, we provide the following models:

File	Domain	Weight Transfer Method	Action Guidance	Training Steps
Vid2World-RT1	RT-1	Extrapolative	✔️	100k
Vid2World-CSGO	CSGO	Extrapolative	✔️	100k
Vid2World-RECON	RECON	Extrapolative	✔️	100k
Vid2World-RT1-NAG	RT-1	Extrapolative	❌	30k
Vid2World-RT1-Masked-NAG	RT-1	Masked	❌	30k
Vid2World-RT1-30k	RT-1	Extrapolative	✔️	30k
Vid2World-RT1-Masked	RT-1	Masked	✔️	30k
Vid2World-RT1-Shift	RT-1	Shift	✔️	30k

Before inference, make sure you switch the |<your_pretrained_checkpoint>| in the config file to the path towards your local checkpoint.

📸 Showcases

🤖 Robot Manipulation 🦾

🎮 Game Simulation 🕹️

🗺️ Open-World Navigation 🧭

For more showcases, check out our Project Page.

🤖 Vid2World for Robot Manipulation

1. Prepare Data & Model

Data

To download and preprocess the used dataset:

• Download the RT-1 Robot Action Dataset from OXE.
• Run the following command in the repo to save the processed dataset to your desired local folder.

python lvdm/data/oxe_data_converter.py --dataset_name fractal20220817_data --input_path {path to downloaded OXE} --output_path {path to stored npz}

Model

For inference, download our corresponding pretrained model from 🤗Huggingface, check out QuickStart.

2. Training

To launch training with the RT-1 dataset, go to configs/manipulation/config_rt1_train.yaml and change the |<your_data_dir>| into the directory where your local data directory. To launch training on 1x4 GPU cards, use the following command:

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base configs/manipulation/config_rt1_train.yaml --train  --name training_512_v1.0 --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

For ablation experiments, we provide the corresponding configurations in configs/ablation.

File	Weight Transfer Method	Action Guidance	Model Checkpoint
config_rt1_*_masked_nag.yaml	Masked	❌	🤗Vid2World-RT1-Masked-NAG
config_rt1_*_extrp_nag.yaml	Extrapolative	❌	🤗Vid2World-RT1-NAG
config_rt1_*_shift.yaml	Shift	✔️	🤗Vid2World-RT1-Shift
config_rt1_*_masked.yaml	Masked	✔️	🤗Vid2World-RT1-Masked
config_rt1_*_all.yaml	Extrapolative	✔️	🤗Vid2World-RT1-30k

3. Inference

Here we provide two setups, one is generating the sequence frame by frame, which is referred to as Auto-Regressive Generation, and one that generates the full sequence all in one go, which we refer to as Non-Auto-Regressive Generation.

Before running the experiments, make sure you download/train the corresponding checkpoints, as well as change the data paths in the config file used.

Auto-Regressive Generation

For auto-regressive generation, run:

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base code_release_configs/manipulation/config_rt1_test_ar.yaml --val  --name training_512_v1.0 --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

While doing ablation, switch the configuration file to the corresponding file.

Non-Auto-Regressive Generation

For non-auto-regressive generation, run:

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base code_release_configs/manipulation/config_rt1_test_nar.yaml --val  --name training_512_v1.0 --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

RT-1 Action Control Test

Test model's ability to respond to different world_vector actions (X+, X-, Y+, Y-, Z+, Z-).

First, update the config file configs/manipulation/config_rt1_action_control_test.yaml:

• Set pretrained_checkpoint to your checkpoint path
• Set data_dir to your RT-1 data directory

Then run:

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base configs/manipulation/config_rt1_action_control_test.yaml --val --name rt1_action_control_test --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

Results will be saved to the directory specified in the config file's save_dir parameter. Each batch visualizes 8 action variants side-by-side for comparison.

🕹️ Vid2World for Game Simulation

1. Prepare Data & Model

Data

To download and preprocess data, please follow the steps from DIAMOND, specifically:

• Download the .tar files in the dataset_dm_scraped_dust2_tars from this dataset repo.
• Use the provided script to process the dataset for full and low res. For our purpose, we use only the full_res folder.

Model

For inference, download our corresponding pretrained model from 🤗Huggingface, check out QuickStart.

2. Training

To launch training with the csgo dataset, go to configs/game/config_csgo_train.yaml and change the |<your_data_dir>| into the directory where your local data directory. To launch training on 1x4 GPU cards, use the following command:

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base configs/game/config_csgo_train.yaml --train  --name training_512_v1.0 --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

3. Inference

Standard Inference

For inference, run:

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base configs/game/config_csgo_test.yaml --val  --name training_512_v1.0 --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

Long Rollout Inference on CSGO

For long rollout inference on CSGO, run:

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base configs/game/config_csgo_test_long_rollout.yaml --val  --name training_512_v1.0 --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

Long Rollout Inference on OOD Games

For long rollout inference on previously unseen games (Valorant, Delta Force), run:

Valorant:

python3 -m torch.distributed.launch --nproc_per_node=2 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base configs/game/config_csgo_test_long_rollout_valorant.yaml --val  --name training_512_v1.0 --logdir |<your_log_dir>| --devices 2 lightning.trainer.num_nodes=1

Delta Force:

python3 -m torch.distributed.launch --nproc_per_node=2 --nnodes=1 --master_addr=127.0.0.1 --master_port=12879 --node_rank=0 ./main/trainer.py --base configs/game/config_csgo_test_long_rollout_delta_force.yaml --val  --name training_512_v1.0 --logdir |<your_log_dir>| --devices 2 lightning.trainer.num_nodes=1

🗺️ Vid2World for Open-World Navigation

1. Prepare Data & Model

Data

To download and preprocess data, please follow the steps from NoMaD, specifically:

• Download the RECON dataset.
• Change the preprocessing resolution to (640,480).
• Run process_recon.py to save the processed dataset to your desired local folder.

Model

For inference, download our corresponding pretrained model from 🤗Huggingface, check out QuickStart.

2. Training

To launch training with the RECON dataset, go to configs/navigation/config_recon_train.yaml and change the |<your_data_dir>| into the directory where your local data directory. To launch training on 1x4 GPU cards, use the following command:

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base configs/navigation/config_recon_train.yaml --train --name training_512_v1.0 --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

3. Inference

Following NWM, we evaluate our performance under two setups: single-step generation and auto-regressive generation. While in both setups, our model is doing auto-regressive generation, the data split is different, we support both setups.

Single-Step Generation

Change the |<data_dir>| and |<path_to_pretrained_checkpoint>| in configs/navigation/config_recon_test_single_step.yaml.

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base configs/navigation/config_recon_test_single_step.yaml --val --name training_512_v1.0 --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

Auto-Regressive Generation

Change the |<data_dir>| and |<path_to_pretrained_checkpoint>| in configs/navigation/config_recon_test_rollout.yaml.

python3 -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --master_addr=127.0.0.1 --master_port=12869 --node_rank=0 ./main/trainer.py --base configs/navigation/config_recon_test_rollout.yaml --val --name training_512_v1.0 --logdir |<your_log_dir>| --devices 4 lightning.trainer.num_nodes=1

🧪 Evaluation

For evaluation, after running the inference code, calculate the metrics by running:

python eval.py --exp_folder |<your_log_image_dir>| --env  |<rt1/csgo/recon_time/recon_rollout>|

📜 Citation

If you find our code useful, please consider citing our paper:

@article{huang2025vid2world0,
  title={Vid2World: Crafting Video Diffusion Models to Interactive World Models}, 
    author={Siqiao Huang and Jialong Wu and Qixing Zhou and Shangchen Miao and Mingsheng Long},
    year={2025},
  journal= {arXiv preprint arXiv:2505.14357}
}

📬 Contact

If you have any questions, please contact huang-sq23@mails.tsinghua.edu.cn.

💡 Acknowledgement

We sincerely appreciate the following github repos for their valuable codebase we build upon:

• https://github.com/Doubiiu/DynamiCrafter
• https://github.com/thuml/iVideoGPT
• https://github.com/facebookresearch/nwm
• https://github.com/eloialonso/diamond
• https://github.com/universome/stylegan-v