Original title: SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models
Junchao Huang, Guian Fang, Shengju Qian, Xianghao Kong +14 · September 2026
Teaching computers to imagine how a visual world changes over time is difficult. Video collections use wildly different formats, visual styles, camera views, and descriptions, while video-generating AI models use different internal designs. Simply tossing these mismatched ingredients together creates confusing training signals and makes experiments hard to repeat. To solve this, researchers built SolarWM. First, a data engine standardizes varied video clips so camera motion, text, and visual quality align cleanly under one set of rules. Next, a three-stage training method trains the models on short clips lasting just 5s. Remarkably, this training allows the resulting systems to run in real time and generate continuous, interactive video for minutes to hours. The creators released their data, code, steps, and model weights so anyone can freely study and build on them.
Imagine hosting a potluck where everyone brings a dish, but recipes are in different languages, measurements are in different units, and cooking methods vary. It's nearly impossible to coordinate the meal without spending all your time converting measurements and adjusting cooking times. In this paper, the researchers create a universal recipe book and kitchen setup that standardizes how dishes are prepared, so all contributions fit together seamlessly for a cohesive meal.
This standardizes how video AI is built, so researchers spend less time fixing mismatched data and more time creating useful simulations for gaming, virtual reality, or training autonomous systems.
The abstract doesn't specify limitations; it presents SolarWM as an open foundation for research, implying it may need further development for specific use cases.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models — Junchao Huang, Guian Fang, Shengju Qian et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.