🌏 LiveWorld: Simulating Out-of-Sight Dynamics in Generative Video World Models

Zicheng Duan^1*  Jiatong Xia^1*  Zeyu Zhang^2*  Wenbo Zhang¹  Gengze Zhou¹ 
Chenhui Gou³  Yefei He⁴  Feng Chen^1†  Xinyu Zhang^5‡  Lingqiao Liu^1†‡

¹Adelaide University  ²The Australian National University  ³Monash University  ⁴Zhejiang University  ⁵University of Auckland

_{* Equal contribution † Project lead ‡ Corresponding author}

LiveWorld enables persistent out-of-sight dynamics. Instead of freezing unobserved regions, our framework explicitly decouples world evolution from observation rendering. We register stationary Monitors to autonomously fast-forward the temporal progression of active entities in the background. As the observer explores the scene along the target trajectory, our state-aware renderer projects the continuously evolved world states to synthesize the final observation, ensuring that dynamic events progress naturally even when entities are completely out of the observer's view.

• Inference code and pretrained weights (initial version in the paper)
• Training code
• Training data preparation pipeline
• Inference sample preparation pipeline
• Demo data and examples (inference + training)
• LiveBench and its inference scripts
• Model trained on 1.3B backbone
• Model (14B/1.3B) with better training

Tested environment:

• Ubuntu 22.04, Python 3.11, CUDA 12.8, PyTorch 2.8 (cu128)
• H100 96GB / H200 140GB, 128GB+ system memory
• Flash Attention 2.8.3 (Flash Attention 3 is also supported on Hopper GPUs but requires building from source)

1. Create conda environment

conda create -n liveworld python=3.11 -y
conda activate liveworld

2. Clone the repository and install ffmpeg

git clone https://github.com/ZichengDuan/LiveWorld.git
cd LiveWorld

# Install ffmpeg (either way works)
conda install -c conda-forge ffmpeg -y
# or: sudo apt-get update && sudo apt-get install -y ffmpeg

3. Install PyTorch

pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128

4. Setup environment (install dependencies + download model weights)

bash setup_env.sh

This script will:

• Install Python dependencies from setup/requirements.txt
• Install LiveWorld and local packages (SAM3, Stream3R)
• Download all pretrained weights (~100GB) into ckpts/
• Download example data into examples/

Place source images in examples/inference_sample/raw/, then run:

bash create_infer_sample.sh

This generates per-image inference configs under examples/inference_sample/processed/, including:

• Scene point cloud (via Stream3R)
• Camera trajectories
• Entity detection and storyline (via Qwen3-VL)

A pre-built sample (kid_coffee) is already included under examples/inference_sample/processed/ — you can skip this step and go directly to inference.

Edit infer.sh to set your config path and GPU, then run:

bash infer.sh

Example infer.sh:

export CUDA_VISIBLE_DEVICES=0
python scripts/infer.py \
    --config examples/inference_sample/processed/kid_coffee/infer_scripts/case1_right.yaml \
    --system-config configs/infer_system_config_few_step_14B.yaml \
    --output-root outputs \
    --device cuda:0

Two system configs are provided:

• configs/infer_system_config_14B.yaml — full-step inference
• configs/infer_system_config_few_step_14B.yaml — 4-step distilled inference (faster)

Place source videos in examples/training_sample/raw/ (organized by dataset name), then run:

bash create_train_sample.sh

This runs a 4-step pipeline:

1. Build samples — clip extraction, entity detection (Qwen3-VL), segmentation (SAM3), geometry estimation (Stream3R), sample construction
2. Captioning — generate text descriptions with Qwen3-VL
3. VAE encode — encode videos to latent space
4. Pack LMDB — package into sharded LMDB for training

Example training samples from MIRA, RealEstate10K, and SpatiaVID_HQ are included under examples/training_sample/.

Edit train.sh to set your GPU configuration, then run:

bash train.sh

Training configs:

• configs/train_liveworld_14B.yaml — 14B backbone
• configs/train_liveworld_1-3B.yaml — 1.3B backbone

Both train.sh and create_train_sample.sh support multi-node multi-GPU — edit the NODES and CUDA_VISIBLE_DEVICES_LIST arrays at the top of each script.

LiveWorld/
├── infer.sh                        # Inference entry point
├── train.sh                        # Training entry point
├── create_infer_sample.sh          # Inference sample preparation
├── create_train_sample.sh          # Training data preparation
├── setup_env.sh                    # One-click environment setup
├── setup/                          # Installation scripts & requirements
├── configs/
│   ├── infer_system_config_14B.yaml        # Full-step inference config
│   ├── infer_system_config_few_step_14B.yaml  # 4-step distilled inference config
│   ├── train_liveworld_14B.yaml            # 14B training config
│   ├── train_liveworld_1-3B.yaml           # 1.3B training config
│   └── data_preparation.yaml               # Data preparation config
├── liveworld/                      # Core package
│   ├── trainer.py                  # Task definition + training loop
│   ├── wrapper.py                  # Model wrappers (VAE, text encoder, State Adapter)
│   ├── dataset.py                  # LMDB dataset loader
│   ├── utils.py                    # Utilities
│   ├── geometry_utils.py           # Geometry & projection utilities
│   └── pipelines/
│       ├── pipeline_unified_backbone.py    # Unified Backbone
│       ├── pointcloud_updater.py           # Stream3R point cloud handler
│       └── monitor_centric/                # Monitor-Centric Evolution Pipeline
├── scripts/
│   ├── infer.py                    # Inference script
│   ├── train.py                    # Training script
│   ├── create_infer_sample/        # Inference sample creation
│   │   ├── assemble_event_bench.py     # Main assembly (trajectory + storyline)
│   │   ├── build_scene_pointcloud.py   # Scene 3D reconstruction
│   │   └── plot_trajectories_3d.py     # Trajectory visualization
│   ├── create_train_data/          # Training data processing steps
│   │   ├── step1_build_samples.py      # Clip extraction + geometry + samples
│   │   ├── step2_captioning.py         # Video captioning
│   │   ├── step3_vae_encode.py         # VAE encoding
│   │   ├── step4a_pack_lmdb.py         # LMDB packing
│   │   └── step4b_cache_keys.py        # Key caching
│   └── dataset_preparation/        # Legacy data preparation
├── examples/                       # Sample data
│   ├── inference_sample/           # Inference example
│   │   ├── raw/                        # Source images (input)
│   │   └── processed/                  # Generated configs + point clouds (output)
│   └── training_sample/            # Training example
│       ├── raw/                        # Source videos (input)
│       ├── processed/                  # Extracted samples (intermediate)
│       └── processed_lmdb/             # Packed LMDB (output)
├── misc/
│   ├── sam3/                       # SAM3 (local package)
│   └── STream3R/                   # Stream3R (local package)
└── ckpts/                          # Model weights (downloaded by setup_env.sh)

iter_input:
  '0':
    scene_text: The brick wall backdrop remains visible behind the stall...
    fg_text: 'On the right, a wooden bench under the wall sits a lovely corgi dog, staying steadily on the bench and rest.'

We thank the authors of Wan2.1, SAM3, STream3R, Qwen3-VL, and DINOv3 for their outstanding open-source contributions.

We also acknowledge that the concept of reasoning about out-of-sight world shares a similar spirit with Out of Sight, Not Out of Mind (Plizzari et al., 3DV 2025) [paper, code], which explored spatial cognition of off-screen objects in egocentric video perception — a totally different domain but a kindred high-level insight.

If you find this work helpful, please consider citing:

@article{duan2026liveworld,
  title={LiveWorld: Simulating Out-of-Sight Dynamics in Generative Video World Models},
  author={Duan, Zicheng and Xia, Jiatong and Zhang, Zeyu and Zhang, Wenbo and Zhou, Gengze and Gou, Chenhui and He, Yefei and Chen, Feng and Zhang, Xinyu and Liu, Lingqiao},
  journal={arXiv preprint arXiv:2603.07145},
  year={2026}
}