SolarWM-Data

The public repository is available on Hugging Face and ModelScope. After download, the local release root is:

SolarWM-Data/releases-v1/

The repository contains portable controls, licenses, small format examples, and the SolarWM-Data-Annotation/ reconstruction package. The full raw-wds/ and latent-wds/ payloads will be distributed separately. See Dataset access before launching a full training or inference example.

Contents

The complete raw-wds/ corpus includes samples marked high, xhigh, and rejected, together with the measurements and intermediate fields needed to build a different mixture. Expensive video, camera, motion, quality, and scene processing is therefore independent of the policy that selects samples or assigns source weights.

The release contains 14 datasets. DL3DV has separate 10s and 60s views, while MiraData, Sekai Walking, and SpatialVid each include an additional clean variant.

Source	Total	High	Xhigh	Rejected
abot	30,966	127	30,715	124
dl3dv-10s	120,924	54,528	60,396	6,000
dl3dv-60s	10,077	3,578	6,065	434
mind	533	117	402	14
miradata	140,877	3,683	17,806	119,388
miradata-clean	135,224	12,865	5,740	116,619
multicamvideo	123,117	89,587	5,369	28,161
omniworld	19,632	3,773	13,552	2,307
realcam_vid	45,697	16,154	16,074	13,469
sekai_game	2,550	537	1,410	603
sekai_walking	22,990	6,054	12,976	3,960
sekai_walking-clean	109,248	46,831	33,660	28,757
spatialvid	365,345	100,815	127,180	137,350
spatialvid-clean	298,514	133,149	73,450	91,915
Total	1,425,694	471,798	404,795	549,101

Every row in this table is part of the frozen annotated corpus. The rejected tier is published because selection is metadata over an already-annotated corpus, so a different threshold, metric or source weighting is an index pass rather than another GPU run. The public annotation package contains no videos; it provides the released annotations and tools needed to reconstruct raw-WDS from source media acquired by the user under the applicable upstream terms.

Each source directory contains tiered WebDataset shards and a full meta.jsonl; portable recipe indexes select the training and test rows. The metadata includes source identity, captions, dimensions, frame rate, camera summaries, selection state, rejection reasons, VMAF, UniMatch, DOVER, saturation, scene-cut, and VLM measurements where available.

Latent generation inventory

Generation	Samples
wan22-ti2v5b-153f-480p-v1	688,424
wan22-ti2v5b-153f-720p-v1	688,418
wan22-ti2v5b-957f-480p-v1	82,321
wan22-ti2v5b-957f-720p-v1	82,321
wan22-i2v-a14b-81f-480p-v1	1,564,464
wan22-i2v-a14b-81f-720p-v1	1,564,464
wan22-i2v-a14b-153f-480p-v1	688,424
wan22-i2v-a14b-153f-720p-v1	688,418
wan22-i2v-a14b-957f-480p-v1	82,321
wan22-i2v-a14b-957f-720p-v1	82,321
minimax-h3-158f-768p-nomind-v1	686,841
ltx-153f-h512-w768	688,418
ltx-953f-h512-w768	82,321

The tar member manifests use backend-specific solarwm_* schemas. Tensor payloads keep their native dtype and shape. Each generation will be published as a separate dataset repository; current repository links and planned generations are maintained in the latent-WDS release list.

Camera convention

Camera trajectories and intrinsics are included in each record. The backend readers expose the same first-frame-relative trajectory to the model. Optional model-side transforms such as logd4 belong to the training and checkpoint contract, not to the dataset.

Recipes and test data

The public recipes are under recipes/clean-81f/, recipes/clean-153f/, recipes/clean-158f-h3/, and recipes/clean-957f/. Each route provides a train index, test index, recipe contract, statistics, and an exclusion index when individual samples are excluded.

Training validation and standalone inference read the selected recipe's test-index.jsonl.gz. Set validation.sample_count and validation.selection_seed to choose the same unique rows at every validation step; the backend's generation seed controls noise independently. The standalone test-set/ is a logical view of canonical primary raw shards, so test semantics are preserved without duplicating video bytes.

example/ contains small records derived from the release itself: one raw sample from each annotation tier and one representative of every published latent reader schema. Its combined index is a portable format catalog for inspection and reader checks, not a training recipe: it deliberately mixes raw and backend-specific latent schemas and includes one rejected raw sample. Example shard paths are relative to example/; source archive paths are relative to releases-v1/ and carry portable object identities for direct traceability. Training uses the homogeneous train-index.jsonl.gz under the selected recipe.

Reading a local release

For a downloaded or mounted release, point both the control root and payload root at the same releases-v1/ directory:

data:
  index_root: /path/to/SolarWM-Data/releases-v1
  transport:
    kind: local
    root: /path/to/SolarWM-Data/releases-v1
  train_index: recipes/clean-153f/latent-wds/wan22-ti2v5b-153f-480p-v1/train-index.jsonl.gz

Every index stores a release-relative object key, so the same downloaded tree can live at any local absolute path. The full storage and integrity rules are in the data contract.

release.json describes the complete logical release across its distribution repositories. Read the license and citation registry in licenses/ before using or redistributing source media or derived tensors.