pixl-data

Training data for pixl, a small text-to-pixel-art model trained from scratch.

Config	Rows	What it holds
photos	~84k	Photo pairs: a 256×256 photo (input), a 32×32, 24-colour version of it (target), and a caption (query)
sprites	~5.7k	CC0 game sprites (image) with a short label (label)
scenes	52k	32×32 scenes built from the sprites (image), with a caption written from the placement (caption)

Load

from datasets import load_dataset

photos = load_dataset("ocdbytes/pixl-data", "photos")    # {"train": ..., "test": ...}
sprites = load_dataset("ocdbytes/pixl-data", "sprites")
scenes = load_dataset("ocdbytes/pixl-data", "scenes")

train = load_dataset("ocdbytes/pixl-data", "photos", split="train")  # one split only

Splits

Config	train	test	Kept together in one split
photos	~82.7k	~1.6k (2% of source photos)	All crops of one source photo (same sha256)
sprites	~5.1k	~540 (10% of groups)	The colour variants of one tile, and the frames of one animation
scenes	50,000	2,000	Set by the generator: test scenes use test sprites, or object pairs never seen in train scenes

Near-copies never appear in both splits, so test scores are not inflated by images the model has already seen. A group is in test when the first 8 hex digits of the SHA-256 of its key, as a fraction of 16⁸, are below 0.02 (photos) or 0.1 (sprites). The split is fixed: new data never moves existing rows.

photos

Source	Caption
Open Images V7	A box label or a relationship between two boxes drawn by people, e.g. "dog", "man riding horse", "wooden table". The photo is cropped around the box.
Wikimedia Commons, Openverse	The search query that found the photo, e.g. "arcade cabinet".

Photos were filtered before upload. A photo was removed when:

• CLIP matched it more closely to prompts for nudity or graphic content (gore, dead bodies, injuries) than to neutral prompts;
• it came from a web search and CLIP found it does not show its query;
• it is CC BY but has no author to credit, or its author field is an email address.

Every photo was changed:

1. Cropped to a square.
2. Resized to 256×256: input.
3. Resized to 32×32 and reduced to 24 colours: target.

Column	Meaning
id	Pair id
provider	openimages, wikimedia or openverse:<provider>
source_id	Id of the photo at its source
query	Caption
title, author	From the photo's source page
license, license_url	Licence of the original photo
source_page, original_url	Where the original photo is
input_path, target_path	File paths in the original folder layout
sha256	SHA-256 of the downloaded original
input, target	The two images

sprites

Sprites from CC0 packs by Kenney (1-Bit Pack, Cursor Pixel Pack, Pico-8 City, Tiny Battle, Tiny Farm, Tiny Town), ansimuz (Sunny Land), GrafxKid (Mini Pixel Pack 3) and Buch (A Platformer in the Forest). label_source says where each label comes from: checked by eye (claude-visual), a file name (filename) or a sprite-sheet atlas (atlas).

Column	Meaning
path	File path in the original pack layout
pack, creator, source_url	Which pack the sprite comes from
license	CC0 1.0
width, height	Size in pixels
label, label_source	Label and where it comes from
image	The sprite (PNG, with transparency)

scenes

Synthetic 32×32 scenes, each made from one sprite style: a background tile repeated in every 16 px slot (8 px for Pico-8 City) and one or two objects, each centred in its own slot. Sprites are never scaled. The caption is written from the placement, so it is always right: "a green pine tree left of a wooden barrel on grass", "a yellow bus at the top right on pavement".

Kind (kind)	Split	Scenes	Objects from
train	train	50,000	Train-split sprites
new sprites	test	1,000	Test-split sprites only
new pair	test	1,000	Train-split sprites in label pairs that never appear together in train scenes

Column	Meaning
caption	The scene's caption
kind	See above
style	Sprite pack (or 1-Bit Pack colour family) the scene is made from
background, background_path	Background tile label and path in the sprites config; empty for a transparent scene
objects, sprites, slots	Each object's label, path and row,column slot, |-separated
image	The scene (PNG, transparent where there is no background)

Licences

Every row has its own licence in license (and license_url for photos):

• Sprites: CC0 1.0.
• Scenes: CC0 1.0, as they are made only from the sprites.
• Photos: CC0, public domain, or CC BY 1.0–4.0. All Open Images photos are listed as CC BY 2.0.
• Open Images annotations (the boxes and labels behind the crops and captions): CC BY 4.0, by Google.

To reuse a CC BY photo, credit it with the row's author, title, source_page and license, and say that it was changed (cropped, resized, colours reduced).

Google lists Open Images photos as CC BY 2.0 but does not guarantee each licence. Wikimedia and Openverse licences come from their providers' metadata. Check a row's source page before relying on its licence.

Known issues

• Web captions are still somewhat noisy: in a sample of 30 web photos that passed the CLIP check, 4 did not show their query.
• The nudity and graphic-content filter is automatic and can miss photos. Please report any you find.
• Many photos show real people. Some show brand logos or photographers' watermarks.