Published by Video Runtime · Source check 2026-09-21 · Evidence method · Suggest a correction
EDITORIALLY VERIFIED PRODUCT INFORMATION
About ReWorld
An interactive world model that streams explorable 704 × 1280 video while you drive the camera with keyboard and mouse. A window-split training scheme decouples precise action following from long-horizon memory, and inference pairs a bounded KV cache with a pose-indexed landmark bank so GPU memory stays constant while revisited places retrieve the right past views. The 5B model is distilled to four denoising steps and trained on metric-aligned multi-source data before streaming chunks back to the player.
Key features
Releases training and inference code plus pretrained checkpoints on Hugging Face and ModelScope. The design explicitly targets constant-memory long rollouts with camera revisits rather than short demo loops.
Getting started
Create the documented Python 3.10 and PyTorch CUDA environment, download the ReWorld-5B checkpoint, then run the supplied inference script and steer the camera with keyboard and mouse.
Real-time interaction
A window-split training scheme decouples precise action following from long-horizon memory, and inference pairs a bounded KV cache with a pose-indexed landmark bank so GPU memory stays constant while revisited places retrieve the right past views. The 5B model is distilled to four denoising steps and trained on metric-aligned multi-source data before streaming chunks back to the player.
Pricing
CC BY-NC-SA 4.0 code; self-hosted compute costs extra Repository code is published under CC BY-NC-SA 4.0. Confirm separate model, dataset and dependency terms before reuse. No hosted price is inferred. Budget for the documented GPU, storage and setup requirements.
Limitations
The licence is CC BY-NC-SA 4.0, which blocks commercial reuse and keeps the project out of permissive open-source groupings. Real-time streaming is a maintainer claim; the README does not publish a full latency distribution or the tested GPU tier for the released checkpoint.
EDITORIAL VIEW
Editor's Verdict
ReWorld is included as a source-available implementation relevant to real-time generative video or interactive world systems.
Run the supplied example on representative hardware and review every linked licence before production use.
How It Works
- A window-split training scheme decouples precise action following from long-horizon memory, and inference pairs a bounded KV cache with a pose-indexed landmark bank so GPU memory stays constant while revisited places retrieve the right past views.
- The 5B model is distilled to four denoising steps and trained on metric-aligned multi-source data before streaming chunks back to the player.
What We Like
- Releases training and inference code plus pretrained checkpoints on Hugging Face and ModelScope.
- The design explicitly targets constant-memory long rollouts with camera revisits rather than short demo loops.
Current Limitations
- The licence is CC BY-NC-SA 4.0, which blocks commercial reuse and keeps the project out of permissive open-source groupings.
- Real-time streaming is a maintainer claim; the README does not publish a full latency distribution or the tested GPU tier for the released checkpoint.
Pricing & Access
- Repository code is published under CC BY-NC-SA 4.0. Confirm separate model, dataset and dependency terms before reuse.
- No hosted price is inferred. Budget for the documented GPU, storage and setup requirements.
Verification Summary
- Official repository and licence reviewed on 2026-09-21. Video Runtime has not installed or benchmarked this project.
- Performance and compatibility statements below are attributed to the maintainers and are not site measurements.
Technical notes & official performance
Official Performance Data
- The README states ReWorld streams 704 × 1280 video in real time after four-step distillation. This is an author claim; Video Runtime has not benchmarked it.
Technical Notes
- Training and inference code were released in August 2026 and the checkpoints in September 2026; the paper is on arXiv (2608.23565).
Sources
- Official GitHub repository ↗github · accessed 2026-09-21
- Code licence ↗github · accessed 2026-09-21
- Official project page ↗official-docs · accessed 2026-09-21
Access, stage & mechanism
These are separate source-checked facts, not a live uptime monitor. Unverified fields are left open; transport and persistent sessions do not establish how frames are generated.
- Release stage
- Not verified
- Access mode
- Not verified
- Source availability
- Not verified
- Transport
- Not verified
- Interaction
- Not verified
- Session mode
- Not verified
- Generation mechanism
- Not verified
- Continuity method
- Not verified
- Licence status
- Not verified
- Tasks
- Not classified
CONTINUE EXPLORING
Related projects & models
Selected by research links and category, not by popularity.
Build long-horizon playable video worlds with camera control, prompt switching and explicit spatial memory.
Another frame streaming option in this section.Public repository; self-hosted setup and compatible GPU required
A streaming interactive world-model release focused on 720p generation and long-horizon memory.
Another frame streaming option in this section.Public repository; self-hosted installation and compatible GPU required
Run an action-conditioned interactive world model with open-ended rollout on a single desktop GPU.
Another frame streaming option in this section.Public repository; self-hosted installation and compatible GPU requiredGenerate continuing interactive worlds with bounded external state, mid-rollout prompts and camera control.
Another frame streaming option in this section.Public repository; self-hosted setup and compatible GPU requiredRead how official-source and hands-on records stay separate in our Methodology.