pixel-art-tiny-adaptive
pixel art generator for English and Arabic
2 min readupdated
A lightweight autoregressive model (24.5M parameters) that turns short text prompts in English, Modern Standard Arabic, or Egyptian Arabic into compact 16x16 RGBA sprites. It uses a native 4,097-color palette from the pixel-art-atlas-1.4m dataset and ships as a pure PyTorch checkpoint with its own inference package. It is a great fit for game developers and hobbyists who need fast generation of tiny pixel art assets from text descriptions.
Key features
- Compact architecture: only around 24.5M active parameters.
- Multilingual input: supports English, Modern Standard Arabic (MSA), and Egyptian Arabic.
- Exact output size: generates 16x16 sprites using a fixed color palette.
- Custom inference: runs via its own PyTorch package rather than a standard diffusers pipeline.
What you can do
- Generate sets of tiny game icons, items, or monsters for indie games.
- Build a Telegram bot or a local script for fast pixel art generation on the fly.
- Test autoregressive generative models on non-standard, low-dimensional outputs.
Example
from huggingface_hub import snapshot_download
import sys
repo = snapshot_download("nomeda-lab/pixel-art-tiny-adaptive")
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("تنين أحمر صغير", seed=42).show()Read more
DetailsBenchmarks: best validation loss is 1.7340 at step 8000.
Training: trained on the Angais/pixel-art-atlas-1.4m dataset (sprites generated with Krea 2 Turbo).
Architecture: FFN with widths [128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128].
Format: custom PyTorch checkpoint (model, config.json, tokenizer.json, and palette.npy).




Comments