Murea Labs' InWorld Turns Typed Actions Into Persistent AI Video Stories
Murea Labs open-sourced InWorld, a video model with synchronized audio plus a planning agent that turns typed actions into a continuing filmed story.
Takeaways
- Murea Labs released InWorld, an open-source video generator with synchronized audio plus a world-state agent.
- Murea-Fast weights are on Hugging Face under Apache 2.0, derived from LTX-2.5 with 8-step distilled sampling.
- Generates 1280x704 24 fps clips with dialogue; 5 seconds in about 17 seconds on four H20 GPUs.
- Murea-Mind agent plans beats, checks frames, and remembers characters, objects, and places across clips.
- Co-op mode supports 2 to 4 players sharing one world, filmed as a single clip per round.
- Murea-Pro scores 81.13 on WBench Full-289; Murea-Fast scores 64.98 on OpenS2V-Eval.
Murea Labs has released the InWorld repository, an open-source stack that converts typed actions into video clips with synchronized sound and dialogue. It combines Murea-Fast, an eight-step video generator distilled from LTX-2.5, with Murea-Mind, an agent that tracks characters, objects, locations, and plot across clips.
The Apache 2.0 model weights are available on Hugging Face. The repository includes a web interface, an HTTP API, and a cooperative mode for two to four participants.
Developers create a scene by supplying ordered reference images for people, objects, and a location, then describing the action in a prompt. InWorld generates 1280×704 video at 24 frames per second. Subsequent clips can inherit the final 17 frames of the previous shot, preserving visual context across cuts.
Two services, one directing loop
| Component | Responsibility | Runtime |
|---|---|---|
| Murea-Fast | Generates video and synchronized audio from prompts and reference images. | Linux, NVIDIA GPUs, LTX-2.5 base components, and a version-pinned SGLang environment. |
| Murea-Mind | Plans shots, validates actions, stores world state, reviews generated frames, and prepares the next prompt. | Python 3.12, uv, FastAPI, React, and an external model provider. |
Murea-Mind supports Gemini, OpenAI-compatible endpoints, and DashScope for language and image tasks. It sends the resulting prompts and visual references to Murea-Fast through the generation service’s HTTP API.
Murea Labs reports that a five-second clip takes about 17 seconds on four H20 GPUs using sequence parallelism, which distributes sequence computation across the devices. Actual latency will depend on GPU type, count, memory, configuration, and workload.