Harvested training shards (shards-v1)

Corpus: 7,514 shards

Each shard holds up to 16 CFA tiles of 256×256 raw sensor pixels. 2,513 shards come from the raw.pixls.us and PlayRaw archives, 5,001 from 12 community contributors. Updated 2026-07-31.

Clean-tile training shards for the Ansel neural raw denoiser — a permanent public cache so nobody repeats the multi-day harvest. The dataset is reproducible from ledger.jsonl plus the harvest scripts in this repository; fetch with scripts/fetch_shards.sh.

Each .npz shard holds up to 16 tiles of 256×256 raw sensor values extracted from one photograph — pristine pre-white-balance ADUs, exactly as read from the sensor (white balance is never applied; the as-shot WB coefficients are merely recorded as metadata) — with sensor calibration (CFA pattern, black/white levels), camera model, ISO and provenance metadata. Shards contain no GPS, timestamps, serial numbers or contributor file paths — but the tiles are viewable fragments of the source photographs.

License and legal use

This corpus is an aggregate of three provenance classes. Every shard's class and license are identifiable from its embedded metadata and the bookkeeping files:

provenance	how to identify	license
raw.pixls.us archive	annex_key field set; ledger.jsonl	CC0-1.0 (the archive is CC0)
PlayRaw community photographs	license, author and topic URL recorded in shard and ledger	as declared per shard: cc0, by, by-sa, by-nc or by-nc-sa, with attribution data embedded
direct contributions	filename prefixed <handle>_; listed in contrib/registry.jsonl	Ansel Training Data License 1.1 (ATDL-1.1, also attached to this release)

Intended use — the one this release exists for: training, validating and testing image-denoising neural networks for the Ansel workflow, on your own infrastructure, as documented in the design page. This use is consistent with every class at once: it is non-commercial free-software development (BY-NC terms), attribution is preserved through the embedded author/source metadata and the ledger (BY terms), and it is the sole use the ATDL-1.0 permits.

What ATDL-1.1 adds for the contributed shards: anyone may also use them, with the ansel-denoise training stack, to audit/reproduce/benchmark the training or to train their own denoising models — whose weights are unrestricted, commercial use included. The bright line is the training stack's capability: the tiles must never feed a stack able to learn anything else than denoising.

Any other use is not covered by the corpus as a whole. In particular: incorporating the tiles into other datasets, "style" learning, training or evaluating generative models, identifying depicted persons or places, or republishing tiles as images. ATDL shards forbid these outright; BY-NC shards exclude commercial use; BY/BY-SA shards impose attribution and share-alike. If you believe a use beyond the intended one is legitimate for a subset, you must filter shard-by-shard on the recorded license, comply with each license individually, and mind that the <handle>_-prefixed shards stay under the ATDL bright line.

Trained weights: the project's position is that network weights trained on these tiles neither contain nor reproduce the photographs and are not distributed as an adaptation of them; released weights carry the Ansel project's own license terms.

Removal: to have your photographs removed, open an issue on this repository. The matching shards are deleted from this release and excluded from all future trainings; weights trained before the removal are unaffected (ATDL-1.1 §5).

No warranty. The data is provided as-is, with no fitness guarantee; neither the contributors nor the Ansel project are liable for uses of it.

Contributing your own images: see CONTRIBUTING.md.