🎬 Demo

Text-driven storytelling with Murea-InWorld · 2 min 19 s · Download the original-quality video

https://github.com/user-attachments/assets/2617c0f3-ded2-4a8f-81bb-b77095c92b0e

[!NOTE] The Murea-Fast weights are available on Hugging Face. Download murea-fast.safetensors and follow the deployment guide to verify the checkpoint and set up the LTX-2.5 base components.

📄 Abstract

Adapted from our technical report. The report evaluates Murea-Pro and Murea-Fast; this repository provides Murea-Mind and the Murea-Fast inference service.

Interactive world models aim to simulate environments that evolve under user actions while retaining the consequences of past interactions. Extending such worlds to sight and sound requires persistent entities, changing scene states, and local audiovisual dynamics to remain coherent across turns. This paper presents Murea-InWorld, an interactive audiovisual world model that couples explicit world-state management with neural audiovisual generation. Its world agent, Murea-Mind, maintains characters, objects, locations, and narrative progress, interprets user actions in context, and compiles planned events into generation conditions. Observations of generated frames inform subsequent planning and selected state updates, closing the interaction loop.

Its audiovisual generator, Murea, realizes these events through Omni-Reference Generation, using visual references to anchor entity appearance and structured descriptions to specify actions and dialogue. Streaming generation uses modality-specific histories, with longer audio context for unfinished utterances. Degraded-history training prepares the model for imperfect rollout context, while sequential MeanFlow and distribution matching distillation enable few-step synthesis. Branching histories support alternative continuations, and co-op interaction brings multiple participants' actions into the same world.

We evaluate the generation components in complementary configurations. Murea-Pro achieves an Overall score of 81.13 on WBench Full-289, the highest among the compared systems. Murea-Fast achieves a Total score of 64.98 on OpenS2V-Eval and an aggregate score of 69.9 on OS2V-Pro, an extended evaluation of reference fidelity, attribute binding, rendering quality, and audio. Murea-Pro and Murea-Fast achieve state-of-the-art performance on their respective benchmarks, and both models surpass most evaluated commercial systems in their respective evaluations.

📣 News

• 2026-10-17 · Watch the text-driven storytelling demo.
• 2026-10-17 · Murea-Fast weights are now available on Hugging Face.
• 2026-10 · First open-source release: the Murea-Fast inference service and Murea-Mind, with documentation in English and Chinese.

✨ Highlights

• Video with sound from several references. Bind people, objects and a place as ordered reference images, refer to them in the prompt, and get 1280×704, 24 fps video with synchronized audio, spoken lines included.
• Clips that keep going. A continuation picks up from the last 17 frames of a clip, with the same people and the same scene.
• Quick to film. 8-step distilled sampling: a 5-second clip takes about 17 s on four H20 GPUs with sequence parallelism.
• An agent that remembers in pictures. Every character, object and place is a reference image with a locked description, so a story keeps its cast and its places from clip to clip. Each beat is planned, checked, filmed and then reviewed against the real frames, and every step is visible in the web UI.
• Play together. In co-op rooms 2–4 players, people or AI teammates, each bring a character, and every round is filmed as one shared clip.

🎮 What you can do

🧩 How it fits together

flowchart LR
    P([Player]) -->|action| A[Murea-Mind<br/>plans, checks, remembers]
    A <-->|text and images| L[(Language and image model APIs<br/>Gemini / OpenAI-compatible / DashScope)]
    A -->|prompt + reference images| F[Murea-Fast<br/>inference service on GPUs]
    F -->|video with sound| A
    A -->|clip + next choices| P

The two parts talk over a documented HTTP API, so the service can run on another machine or behind an SSH tunnel, and other clients can use it too.

🧭 Where to start

⚙️ Components

🚀 Quick start

1. Start Murea-Fast on a GPU machine. Install the environment and the LTX-2.5 base components as described in Deployment, download the murea-fast.safetensors checkpoint, and set the four paths in inference/configs/single-gpu.json. Set checkpoint to the absolute path of the downloaded file. Then:

cd inference
bash scripts/murea-fast prepare --config configs/single-gpu.json
bash scripts/murea-fast start --config configs/single-gpu.json
curl --fail http://127.0.0.1:30260/health

2. Run Murea-Mind on any machine. It needs uv and a key for text and images (Gemini, any OpenAI-compatible server, or DashScope):

git clone https://github.com/Murea-Labs/InWorld.git Murea-InWorld && cd Murea-InWorld/agent
cp .env.example .env          # set GEMINI_API_KEY and MUREA_URL=http://<gpu-host>:30260
./run.sh                      # then open http://127.0.0.1:8080

See the Murea-Mind README for other providers, Docker and SSH tunnels.

🗂️ Repository layout

Murea-InWorld/
├── LICENSE
├── CONTRIBUTING.md
├── SECURITY.md
├── inference/            Murea-Fast inference service
│   ├── configs/          service configurations and an example request
│   ├── docs/             deployment, API, performance, provenance
│   ├── scripts/          CLI wrapper, connector merge, acceptance run
│   ├── src/murea_fast/   launcher and client
│   ├── tests/
│   └── vendor/sglang/    pinned SGLang with the model implementation
└── agent/                Murea-Mind
    ├── murea_agent/      the agent (Python package)
    ├── app/              FastAPI server
    ├── frontend/         React UI (prebuilt in frontend/dist)
    ├── docs/             co-op mode design notes
    └── tests/

🔒 Security

Neither the inference service nor the agent has authentication. Read SECURITY.md before deploying them on a shared network.

🤝 Contributing

See CONTRIBUTING.md for the checks each part needs before a pull request.

📜 License

The code is released under the Apache License 2.0. The vendored SGLang in inference/vendor/ keeps its own copyright and license notices, including the files it carries under MIT and BSD-3-Clause. The Murea-Fast model weights are a derivative of LTX-2.5 and are governed by the LTX-2.x Community License Agreement; see Provenance for details. The agent calls third-party language and image model APIs, which have their own terms.

📚 Citation

@misc{murea_inworld2026,
  title        = {InWorld: From Imagination to Living Worlds with Agents and Fast Video Generation},
  author       = {{Murea Labs}},
  year         = {2026},
  url          = {https://murea.ai/page/paper/InWorld-technical-report.pdf}
}

✉️ Contact

Email: labs@murea.ai