Long-horizon interactive video world models fail mostly because training corpora are fragmented by source, temporal scale, camera geometry, and captioning style. SolarWM-Data attacks that upstream bottleneck by converting many heterogeneous video sources into a single, camera-conditioned, frame-aligned corpus and by separating expensive source preprocessing from mixture construction, making large-scale, reproducible world-model training practical.
What Sets It Apart
- Unified canonical contract: each clip aligns raw frames, metric camera geometry, captions, quality metadata, selection decisions, and provenance — so models get consistent, camera-conditioned supervision across sources rather than noisy ad-hoc mixes.
- Multi-source scale and portability: the release provides 1,425,694 canonical clips (reported corpus), with the processed dataset hosted as reader-ready shards (approx. 2.73 TB). Raw sources correspond to a much larger physical corpus, and rejected shards and selection metadata are preserved for reproducible filtering.
- Backbone-native generations and latent payloads: published latent-wds and reader-ready generations for multiple backbones (examples include Wan 2.2 variants, MiniMax-H3, and LTX-2.5) so researchers can train or evaluate while preserving backbone representations and objectives.
- Reproducibility & configurability: the repo includes annotations, deterministic test indexes, reconstruction tools, and recipes so you can rebuild raw webdataset shards, reweight sources, or define alternate mixtures without re-running source-level preprocessing.
Who It's For and Trade-offs
Great fit if you are training or benchmarking camera-conditioned, long-horizon video world models and need a single, reconfigurable multi-source corpus with provenance and camera metadata. It’s also useful if you want backbone-native latent generations for method comparisons. Look elsewhere if you need a small, instantly streamable dataset (this release is large to download) or if your project cannot accommodate mixed license constraints: code is Apache-2.0 but some bundled backbone assets use their own community licenses and some raw sources must be obtained under original terms.
