OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Liveworld

Liveworld is an open-source Python research implementation of a generative video world model that separates scene evolution from camera observation. It advances active entities outside the viewer’s field of view and renders observations from their updated states. The repository provides inference and training code, data-preparation pipelines, examples and pretrained weights.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Liveworld — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Liveworld to explore or build on research workflows for video generation where scene events should continue beyond the camera’s view, or to prepare data and train the provided model.

Who it is for

  • Researchers studying generative video world models and persistent off-screen scene dynamics.
  • Computer-vision and video-generation practitioners with the GPU resources to run the provided inference or training workflows.

Documented uses

  • Generate video observations along a camera trajectory while allowing unseen entities to continue changing.
  • Prepare image-based inference configurations, including scene geometry, camera trajectories, detections and storylines.
  • Convert source videos into training samples through detection, segmentation, geometry estimation, captioning, encoding and LMDB packaging.
Before you choose

Requirements & limits

The project is a public GitHub repository under the Apache-2.0 license. The documented setup downloads approximately 100GB of pretrained weights and example data.

  • The implementation’s central design separates world evolution from observation rendering, advancing active entities through stationary monitors while the observer explores.
  • Inference preparation accepts source images and constructs per-image configurations with a point cloud, camera trajectories, entity detections and a storyline.
  • Training preparation accepts source videos and packages processed samples into sharded LMDB.
  • The README documents a high-resource environment: Ubuntu 22.04, Python 3.11, CUDA 12.8, PyTorch 2.8, H100 96GB or H200 140GB GPUs, and at least 128GB system memory.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Prepare and run an inference sample

Generate a Liveworld video inference output from source images or the included prepared example.

Before you start

  • Ubuntu 22.04, Python 3.11, CUDA 12.8 and PyTorch 2.8.
  • The README lists H100 96GB or H200 140GB GPUs and 128GB or more system memory.
  • Install ffmpeg and run the repository setup script; setup downloads approximately 100GB of pretrained weights and example data.

Steps

  1. Install the documented environment and run the command below from the repository.

    Documented command
    bash setup_env.sh

  2. For your own images, place them in examples/inference_sample/raw/ and run the command below. Alternatively, use the included kid_coffee processed sample.

    Documented command
    bash create_infer_sample.sh

  3. Edit infer.sh to set the configuration path and GPU.

  4. Run the command below; select the full-step or four-step distilled system configuration as appropriate.

    Documented command
    bash infer.sh

Expected outcome

The inference script uses the selected sample configuration and system configuration to generate output under the configured output root.

Guide limitations

  • The documented environment and GPU-memory requirements are substantial.
  • The setup script downloads approximately 100GB of weights and example data.
Documentation 1 ↗
Source-based practical guide

Prepare video data and launch training

Build training samples from source videos and start training with a supplied configuration.

Before you start

  • The repository’s documented Ubuntu, Python, CUDA and PyTorch environment, plus the listed GPU and system-memory capacity.
  • Source videos organized by dataset name under examples/training_sample/raw/.
  • Pretrained weights and dependencies installed through the repository setup.

Steps

  1. Place source videos in examples/training_sample/raw/.

  2. Run the command below to extract clips, detect entities, segment, estimate geometry, caption, VAE-encode and package samples into sharded LMDB.

    Documented command
    bash create_train_sample.sh

  3. Edit train.sh to set the GPU configuration and choose a supplied training configuration for the 14B or 1.3B backbone.

  4. Run the command below.

    Documented command
    bash train.sh

Expected outcome

The documented pipeline prepares video samples for training and starts training with the selected configuration.

Guide limitations

  • Training requires GPU resources; the README says the preparation and training scripts support multi-node, multi-GPU setups through their node and GPU arrays.
  • Example prepared training samples are included, but this recipe describes the documented path for preparing data from source videos.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
ZichengDuan/LiveWorld · Official homepage title · retained in its source language
How we review sourcesSuggest a correction