AIAny
Icon for item

SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models

Provides an end-to-end, reproducible foundation for camera‑controllable, long‑horizon video world models — converting 1.43M clips from 10 datasets into a unified canonical corpus and releasing data, pipelines, recipes, and weights. Introduces backbone‑native adaptation and a three‑stage training recipe to produce 5B–33B models that enable minute‑to‑hour rollouts after training on 5s sequences.

Introduction

Most progress in video generation focuses on short clips and model‑specific pipelines; building interactive, long‑horizon world models instead requires consistent supervision across diverse datasets and support for heterogeneous video generators. SolarWM's central insight is to decouple heavy source preprocessing from mixture construction via a canonical, frame‑aligned data contract and to adapt heterogeneous backbones through lightweight, backbone‑native interfaces — enabling reproducible cross‑backbone training and long‑horizon causal rollouts without long‑video pretraining.

Key Findings
  • Canonical multi‑source corpus: the pipeline converts ~1.43 million canonical clips from 10 datasets (real, synthetic, and game environments) into a single frame‑aligned representation that includes visual frames, metric camera geometry, captions, quality metadata, selection records, and provenance. So what: researchers can reconfigure mixtures (filters, sampling ratios, source weights) without redoing expensive source preprocessing.
  • Backbone‑native adaptation: SolarWM preserves each video generator’s native representation and objectives, adding only interfaces for camera conditioning and rollout. So what: the same data and training recipe can be applied to diverse backbones (Wan2.2, LTX, MiniMax families) enabling fair cross‑backbone comparison and reuse.
  • Simple three‑stage recipe: bidirectional adaptation (bidirectional training), teacher‑forced autoregressive (AR) initialization, and distribution matching distillation (DMD). So what: this pipeline yields causal, low‑latency generators that support minute‑to‑hour rollouts after being trained only on 5s sequences, avoiding expensive long‑sequence optimization.
  • Scalable results: instantiations span 5B–33B parameters (SolarWM variants based on Wan2.2, LTX‑2.5, MiniMax‑H3) and demonstrate real‑time interactive rollouts across minutes to hours. So what: long‑horizon behavior emerges from proper data construction and staged adaptation rather than massive long‑sequence training.
Who it's for and tradeoffs

Great fit if you are a researcher or engineering team aiming to: reuse diverse video sources for world‑model training, compare or adapt multiple video generator backbones under a shared data contract, or build camera‑controllable, causal video rollouts for interactive applications. SolarWM is especially useful when you need reproducible mixtures and want to avoid reprocessing terabytes of raw sources. Look elsewhere if you require out‑of‑the‑box long‑video fine‑tuning workflows specialized for a single, proprietary backbone, or if you need productionized low‑latency distilled generators (the project releases recipes and weights but final inference latency depends on downstream distillation and serving work).

Where it fits

SolarWM sits between raw multi‑source dataset efforts and single‑backbone world‑model papers: it standardizes the data side (canonical corpus + reconfigurable mixture) and provides a lightweight, backbone‑preserving adaptation and training recipe so that community models and datasets can be evaluated and extended in a consistent, reproducible way.

Information

  • Websitearxiv.org
  • OrganizationsCUHK-SZ SLAI, NUS, CUHK, HKUST, HKUST-GZ, NVIDIA, UCLA, MSRA
  • AuthorsJunchao Huang, Guian Fang, Shengju Qian, Xianghao Kong, Zhuoran Zhao, Wei Huang, Yihua Du, Zixin Zhang, Justin Cui, Yuchao Gu
  • Published date2026/09/02

Categories

More Items

Generates synchronized native 2K audio-video from a single first frame and a text prompt using a compact 7B joint generator. Combines gated cross-modal attention, progressive joint training, audio-video reinforcement learning, and an Autoregressive 1-Step 2K Refinement; releases a 7B generator and 2K Refiner for research use.

Measures whether video generators reproduce the correct distribution of possible physical behaviors under repeated rollouts. Introduces PAWBench and PAWEval to convert repeated generations into outcome-level empirical distributions and quantify probabilistic alignment; evaluates 50 scenarios and 11 models and finds no model consistently matches reference probabilities.

Evaluates visual reasoning in video generation models using 27 photorealistic tasks (810 instances), a two-level taxonomy of domains and skill tags, and task designs that enforce valid intermediate trajectories and calibrated difficulty.