cver91/uspl1

Unsplash JPEG Max

Full-resolution Unsplash photos, fetched from the Unsplash CDN at the highest JPEG quality it serves (?fm=jpg&q=100&chromasub=444: ≈ IJG quality 98, 4:4:4, original pixel size), plus per-photo metadata. Unsplash+ (premium, watermarked) photos are not included.

Images	2,094,223 (18.0TB)
Image shards	600 × ~30 GB (data/train-*.tar)
Platform metadata rows	146,294 in 1 part(s) (platform/part-*.parquet)

Last updated 2026-10-05 17:02 UTC. The dataset is uploaded incrementally; new shards and parts are appended, existing files never change.

Layout

• data/train-XXXXXX.tar — WebDataset shards. Each sample is <photo_id>.jpg (bytes exactly as served by the CDN) + <photo_id>.json (list record, file info incl. sha256 / size / quality, access flags incl. flagged, the original's EXIF from the CDN cdn_meta, and napi platform metadata when it had already been collected at download time).
• metadata/index/train-XXXXXX.parquet — one row per image of the matching shard (photo_id, size, sha256, dimensions, caption, page URL, …).
• platform/part-XXXXXX.parquet — Unsplash platform metadata (description, tags, likes/views/downloads, location, EXIF, photographer, …), latest snapshot per photo; raw_json holds the complete original record. Coverage grows over time and is independent of the image shards; join on photo_id = id.

Loading

from datasets import load_dataset
images   = load_dataset("cver91/unsplash-jpeg-max", split="train", streaming=True)
index    = load_dataset("cver91/unsplash-jpeg-max", "index", split="train")
platform = load_dataset("cver91/unsplash-jpeg-max", "platform", split="train")

License

Photos are distributed on Unsplash under the Unsplash License. Keep photographer attribution (page_url, user_name / user_username) when using them.