license: other
license_name: ltx-2-community
license_link: https://github.com/Lightricks/LTX-2/blob/main/LICENSE-2_x
pipeline_tag: image-to-video
tags:
- video-generation
- audio-video-generation
- multi-reference
- video-continuation
- world-model
- murea
Murea-Fast
From imagination to living worlds with agents and fast video generation.
🎬 Demo · 📦 Checkpoint · 🚀 Get started · 🌏 中文简介 · ✉️ Contact
Murea-InWorld turns actions and imagination into audiovisual stories. Murea-Mind plans each story beat, maintains explicit world memory, and checks generated observations. Murea-Fast brings those plans to life as video with synchronized sound, conditioned on text and ordered visual references.
This repository hosts the released Murea-Fast checkpoint. The inference service, Murea-Mind application, tests, and deployment instructions are available in the code repository. Murea-Pro is evaluated in the technical report; its weights are not included in this release.
🎬 Demo
Watch the storytelling demo · Explore more worlds
✨ Highlights
- Multi-reference audiovisual generation: use ordered references for characters, objects, and locations, then describe actions and dialogue in the prompt.
- Video with synchronized sound: generate audiovisual clips with an 8-step distilled sampler.
- Continuation: extend a clip from a clean latent prefix derived from its last frames.
- Interactive worlds with Murea-Mind: turn player actions into a continuing story with persistent entities, branching histories, and shared co-op adventures.
📦 Released checkpoint
| Item | Details |
|---|---|
| Model | Murea-Fast |
| File | murea-fast.safetensors |
| File size | 42,018,192,440 bytes, approximately 42.02 GB |
| Sampling | 8 steps |
| Validated output setting | 1280 × 704 pixels, 24 fps |
| Example clip | 121 frames, approximately 5 seconds |
| Continuation | 17-frame prefix; new video frames are generated after the prefix |
| Runtime | The Murea-Fast service in the GitHub repository |
SHA-256:
1a2c0d8e23032994b52f2c67e14c930d3be8224f321c51353f0a2954ccbb9385
🚀 Download and run
1. Download the weights
python -m pip install -U huggingface_hub
hf download MureaLabs/InWorld murea-fast.safetensors --local-dir ./weights
2. Get the inference code
git clone https://github.com/Murea-Labs/InWorld.git
cd InWorld/inference
Follow the deployment guide to install the GPU environment and obtain the required tokenizer, text encoder, connectors, audio/video decoders, and vocoder. The checkpoint alone is not a complete inference installation.
Set the absolute python, model_dir, checkpoint, and runtime_dir paths in configs/single-gpu.json. Set checkpoint to the downloaded murea-fast.safetensors file. Choose a configuration appropriate for the available GPUs and memory.
After completing the environment and component setup:
bash scripts/murea-fast verify /absolute/path/to/murea-fast.safetensors --sha256
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
For reference-conditioned generation and continuation, use the inference quick start and HTTP API reference. To play stories in the browser, follow the Murea-Mind quick start.
📊 Evaluation
The technical report reports the following results for Murea-Fast:
| Benchmark | Metric | Score |
|---|---|---|
| OpenS2V-Eval | Total | 64.98 |
| OS2V-Pro | Aggregate | 69.9 |
See the technical report and project results for the evaluation protocols and comparisons.
🖥️ Hardware and practical limits
The released inference service has been validated on NVIDIA H20 GPUs in one-GPU and four-GPU configurations. Runtime and memory requirements depend on reference image sizes, clip length, parallelism, and software settings. Consult the measured performance and memory results before choosing hardware.
- The published checks cover 5-second clips on one H20 and clips up to approximately 20 seconds on four 141 GB H20 GPUs. These results do not establish compatibility with smaller-memory GPUs.
- Continuation regenerates audio; continuous speech and sound across clip boundaries are not guaranteed.
- Multiple subjects can merge or lose identity, and long stories can drift. The available deployment checks do not establish quality across all prompts or worlds.
- The current service does not support frame interpolation or upscaling. Experimental acceleration settings need separate validation.
🌏 中文简介
Murea-InWorld 将文字行动与参考图像转化为持续发展的视听故事：Murea-Mind 负责世界记忆、剧情规划与生成结果检查，Murea-Fast 负责生成带同步声音的视频。
本模型仓库发布的是 Murea-Fast 权重 murea-fast.safetensors；Murea-Mind、推理服务和完整使用文档位于 GitHub 代码仓库。论文中评测的 Murea-Pro 权重不包含在本次发布中。
- 支持有序多参考图条件、8 步蒸馏生成和视频续接。
- 已验证的输出设置为 1280 × 704、24 fps；显存与耗时请参阅实测记录。
- 下载权重后，还需按部署文档安装运行环境和配套组件。
- 续接时音频会重新生成，跨片段的语音和声音连续性不作保证。
中文部署说明 · 中文推理指南 · Murea-Mind 中文说明
📜 License
The InWorld code is released under Apache License 2.0.
The Murea-Fast weights and required base components are governed by the LTX-2.x Community License Agreement, as documented in the repository's provenance and license notice. The code license does not replace the model-weight license. See the linked agreement for its use and redistribution terms.
📚 Citation
@techreport{murealabs2026inworld,
title = {InWorld: From Imagination to Living Worlds with Agents and Fast Video Generation},
author = {{Murea Labs}},
institution = {Murea Labs},
year = {2026},
url = {https://murea.ai/page/paper/InWorld-technical-report.pdf}
}
✉️ Contact
Email: labs@murea.ai
Visit the project website or open an issue in the code repository for questions and reproducible bug reports.