Tiny Arabic/English Pixel-Art Generator

A small conditional autoregressive model that turns short English, Modern Standard Arabic, Egyptian Arabic, or mixed prompts into 16ร—16 RGBA pixel-art sprites, using a compact 513-colour palette (index 0 = transparent) remapped from the dataset's native 4,097 colours.

  • Active parameters: 23,121,344
  • FFN widths: [128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128]
  • Output: 256 compact palette indices (0 = transparent, 1โ€“512 = colour) โ†’ 16ร—16 RGBA
  • This is a custom PyTorch checkpoint, not a diffusers pipeline.
  • Best validation loss: 1.2000 at step 3000

Usage

from huggingface_hub import snapshot_download
import sys

repo = snapshot_download("nomeda-lab/pixel-art-tiny-colorizer")
sys.path.insert(0, repo)

from pixelgen.inference import PixelArtGenerator
gen = PixelArtGenerator.from_pretrained(repo)
gen.generate("a small red dragon", seed=42).show()
gen.generate("ุชู†ูŠู† ุฃุญู…ุฑ ุตุบูŠุฑ", seed=42).show()

config.json, model.pt, tokenizer.json, and palette.npy are the model. The pixelgen/ package and requirements.txt are included.

Data and licence

Trained on Angais/pixel-art-atlas-1.4m (CC0-1.0 with provenance caveats; sprites are AI-generated with Krea 2 Turbo). The pixelgen code is MIT. Check the dataset card before commercial use of generated assets.

Downloads last month
42
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support