Download README.md from DiogenesLab/Diogenes_Factory_Training_Iterations_Archive_v01: direct link, hf CLI and curl.
- Browser
- Download file 5.25 kB
-
https://huggingface.co/datasets/DiogenesLab/Diogenes_Factory_Training_Iterations_Archive_v01/resolve/main/README.md
- Command line
-
hf download hf://datasets/DiogenesLab/Diogenes_Factory_Training_Iterations_Archive_v01/README.md
-
curl -L -o README.md https://huggingface.co/datasets/DiogenesLab/Diogenes_Factory_Training_Iterations_Archive_v01/resolve/main/README.md
pretty_name: Diogenes Factory Training Iterations Archive v01
license: other
license_name: owner-authorized-scene-capture-archive
license_link: >-
https://huggingface.co/datasets/DiogenesLab/Diogenes_Factory_Training_Iterations_Archive_v01/blob/main/NOTICE.md
tags:
- video
- simulation
- multi-view
- camera-pose
- archival
size_categories:
- 1K<n<10K
Diogenes Factory Training Iterations Archive v01
Historical rendered scene captures, synthetic camera plans, action-camera arrays, technical provenance, quality reports and failed capture attempts from the closed Diogenes project. The complete source scope is 4,249 files / 46,676,729,553 bytes, including local iterations and the separate multicamera delivery subset. Iterations and delivery copies can overlap; file counts do not represent independent capture hours or a train/test split.
Archival publication is complete: 4,249 logical files are preserved in 56 TAR shards. All 675 MP4/WAV/PNG source files retain their original bytes. The 3,574 technical files include 576 explicit privacy copies; privately retained originals have full SHA-256 proofs. MANIFEST.jsonl identifies every actual public member and ARCHIVE_INDEX.jsonl records immutable upload revisions and whole-shard SHA-256 values. Original scoped R2 objects and temporary public copies were removed only after pinned Hugging Face LFS verification and anonymous Range checks. These checks are distinct from an anonymous full download of every large TAR.
| Historical type | Files |
|---|---|
.active_frac |
19 |
.json |
2,805 |
.log |
229 |
.md |
17 |
.mp4 |
488 |
.png |
76 |
.wav |
111 |
.xml |
116 |
[status marker] |
388 |
Layout and interpretation
collections/<iteration>/... preserves local capture iteration paths. engine-delivery/<scene>/... preserves the delivery subset as a distinct source collection. A completed clip commonly has four video lanes, action-camera arrays, system information, camera plans, capture provenance and available dynamic-object/co-visibility checks. Some attempts contain only status markers or failed tests.
Camera trajectories and generated player/input fields describe synthetic engine experiments, not recorded human controls. Use each historical field_conventions record and the actual video geometry before comparing axes, rotations, aspect ratios or time bases. Recordings include nominal 60-second and 300-second experiments and landscape/portrait variants. Audio provenance identifies an engine mix captured with Unity AudioRenderer and muxed into the video lanes; this is not a natural-world microphone corpus.
Failed QC, historical internal-validation/evaluation restrictions, void states, non_trainable and production-certification fields remain unchanged. Empty smoke scenes, occluded cameras, incomplete frames and incorrect framing remain research evidence. The Factory Training iteration factory-mc-v2/clip-009903 includes four truncated camera JSON files and at least one MP4 lacking a movie metadata atom; these source bytes are retained without fabricated repairs. This statement describes that Factory iteration, including when reading the paired city archive.
Publication authority and attribution
Factory Training is attributed to Unity Technologies, upstream asset listing. On 2026-11-15 the project owner explicitly confirmed additional dataset-publication authorization and instructed complete public migration of both scene collections. This publication relies on that owner confirmation; the archive process did not independently inspect a newly supplied written license. Historical Asset Store EULA and internal-use tags are preserved as provenance, rather than rewritten into a broader grant.
The archive contains rendered outputs and project-generated technical records, not Unity editor binaries, asset source packages or executables. Public access and the project's Apache-2.0 code license do not relicense upstream assets, grant rights to reconstruct or redistribute them, or establish every downstream model-training use. See NOTICE.md.
Privacy and verification
Representative review across the two scene collections inspected 115 video first frames and 26 original PNG inspection sheets. No private desktop or account overlay was identified in those samples. This was not exhaustive viewing or audio transcription. One additional sampled MP4 could not decode because movie metadata was absent and remains an explicitly incomplete original.
Original MP4, WAV and PNG bytes are preserved. Technical public copies remove personal roots, workstation/network identifiers and Unity licensing/runtime diagnostics while preserving JSON numeric tokens, booleans, array structure and historical QC results. PUBLIC-COPIES.json maps original and public hashes; original private files remain in a SHA-verified audit backup. Fixed-revision Hugging Face payload verification precedes any corresponding R2 source deletion.
Source tooling: Diogenes-Archive. Demo: diogenesmodel.com.