sprite32 / DATASHEET.md
danemarguglio's picture
sprite32 1.0 and 1.0 fast
25c7d1e
|
Raw History Blame Contribute Delete
4.49 kB
# Training data
What sprite32 was trained on. The images are not shipped with the model. `attribution.tsv.gz` has a row
per image and `SOURCES.md` a row per source.
## What is in it
660,585 RGBA images at 32x32 from 270 source folders. 653,980 are used for training and 6,605 (1%) for the
validation loss. The 5,000 held-out benchmark images are separate and not counted here.
| Part | Images |
|---|---:|
| First collection: icon sets, emoji sets, game art, OpenGameArt, itch.io packs, clip art | 462,967 |
| Second collection: open-source game repositories with share-alike or GPL art | 67,303 |
| Second collection: other game art | 21,702 |
| Second collection: vector art | 17,709 |
| Second collection: desktop icon themes | 16,824 |
| Second collection: wider search | 7,742 |
| Second collection: Wikimedia Commons | 6,338 |
| Short strings drawn with bitmap fonts by `scripts/gen_rendered_text.py` (fonts listed in the script) | 60,000 |
By label: icon 291,689; pixel art 162,313; game sprite 111,181; UI element 25,570; emoji 23,872; text 22,941;
tile or texture 14,253; other 8,766. Quality label 5: 154,170; 4: 308,490; 3: 181,760; 2: 16,165.
The largest sources are the Iconify icon sets (176,030 images in two folders), the rendered text (60,000),
OpenGameArt (156,848 over eleven folders by licence), Kenney (50,429) and Openclipart (26,242).
## How it was put together
- Sources were downloaded from their repositories, their own sites or mirrors. Larger images were scaled
down to 32x32 with premultiplied alpha.
- Every image except the rendered text has a short and a long caption in English, a category and a quality
score from 1 to 5, all machine-written, and went through a keep-or-drop pass. Of the 154,635 candidates in the second collection, 137,618 were kept; the drops include 151 brand logos
and 31 sexual images.
- 101,504 images (15.4%) sit in 34,012 clusters of near-duplicates (within RMSE 10 of each other), mostly
animation frames and colour variants. They stay in, but sampling is capped at 150 expected draws per
image and 200 per cluster over the run.
- 124,475 flat black icons are shown in one of 16 other colours half of the time, with the colour word in
the caption changed to match.
- The held-out set was drawn from the first collection, before the second existed. 818 of its 5,000 images
have a training image within RMSE 10 (16 are identical). That is the 16.3% quoted in the model card.
## Licences
| Licence family | Images |
|---|---:|
| Public domain / CC0 (includes the 60,000 generated) | 236,141 |
| Permissive, notice to be kept (MIT, Apache-2.0, ISC, BSD) | 155,562 |
| Attribution (CC BY, OGA-BY) | 149,929 |
| Share-alike or copyleft (CC BY-SA, GPL, LGPL, OFL, Free Art) | 118,466 |
| Other (free-use statements, GFDL, government open licences) | 487 |
| How the licence is known | Images |
|---|---:|
| Verified: read from the upstream licence file or page | 305,895 |
| Verified, with conditions (attribution, share-alike, copyleft, trademark note) | 177,092 |
| Stated: what the source or its mirror says; upstream page not fetched | 117,598 |
| Generated by a script in this repository | 60,000 |
The stated-only images are almost all OpenGameArt entries taken from a Hugging Face mirror that records each
entry's licence, author and URL (115,141 of 117,598).
## Left out
- Images that the two licence audits flagged as problems (for example an icon set whose upstream licence
changed to non-commercial, no-derivatives), and the `icon-dataset` source. 137,246 images of the first
collection are not in the training set; that number also includes the held-out set and an older
validation split.
- Two projects that were taken to opt out of AI training (Aurora.3 and CEV-Eris).
- Images made by another text-to-image model.
## Known gaps
- Licence checking was best effort. It was done per source or per sub-collection, not by examining each
image, and a source can contain files whose real origin differs from what it states.
- Some icon sets contain product logos, and some game art depicts existing characters. The licence of the
drawing does not cover the trademark or the character.
- 63,402 images have no author recorded: 57,955 are CC0, and 5,447 are MIT, Apache-2.0 or ISC icons whose
notice is the icon set's licence file (the set is named in the `source` and `url` columns).
- Captions and labels are machine-written and were spot-checked on contact sheets, not reviewed one by one.