Datasets:
pretty_name: 32x32 Pixel Icon Dataset (Z-Image-Turbo)
tags:
- image
- pixel-art
- synthetic
- diffusion
task_categories:
- unconditional-image-generation
- text-to-image
size_categories:
- 1K<n<10K
32x32 Pixel Icon Dataset
1059 synthetic 32x32 RGB icon images generated with Tongyi-MAI/Z-Image-Turbo (zimage-turbo), built for training a from-scratch diffusion pipeline (normalize -> add noise -> train -> predict noise -> sample).
License: TODO -- set before publishing. Check the license of the generating model (Tongyi-MAI/Z-Image-Turbo) and this repo's own terms for any restrictions that apply to the generated images before choosing a license here.
Generation
Images were rendered at 512px and downsampled to 32px with a progressive Lanczos pipeline (halving each step, never more than 2x per call) followed by a light unsharp mask (sharpen=0.6) to recover edge definition.
Quality filtering:
- degenerate filter: drops flat/blank/over- or under-exposed outputs (min_colors=10, min_lum_std=10)
- perceptual dedup: dhash (dhash_8x9, distance<4) within each (subject, style) group removes near-duplicates
Stats
- prompt combos: 232
- attempted: 1160
- kept: 1059
- filtered (degenerate): 1
- filtered (duplicate): 100
- generation failures: 0
Categories:
- animal: 12 subjects
- vehicle: 8 subjects
- object: 10 subjects
- food: 7 subjects
- nature: 8 subjects
- building: 5 subjects
- character: 8 subjects
Structure
Standard HuggingFace imagefolder layout -- images and metadata.jsonl share
this directory, keyed by file_name:
<file_name>.png
metadata.jsonl # file_name, prompt, category, subject, style, seed, gen_size, dhash, metrics
manifest.json # dataset-level generation params (not read by the loader)
Load with:
from datasets import load_dataset
ds = load_dataset("imagefolder", data_dir=".")
Each metadata.jsonl record also carries per-image metrics (colors,
luminance std, edge density) captured at generation time.