UbaNova-16

Text to 16×16 pixel-art game sprites. UbaNova-16 is a small (42.4M-parameter) autoregressive transformer that draws a sprite one pixel at a time from an English prompt. Every pixel is an exact palette color and the background is truly transparent, so the output is clean pixel art with no blur and no post-processing.

Created by Muhammad Fahril Irkham (@UbaNova) · GitHub · Hugging Face

📄 Paper: doi:10.5281/zenodo.24214385 (Zenodo) · English PDF · Bahasa Indonesia PDF

Four samples each for "polished bladed mace", "tiny double cheeseburger", "legendary iron hauberk with gold trim", "ice pop with frost crystals", "futuristic pink digital camera with a green screen", "eagle nest", "silver magic staff", "red microphone with a brass trim".

Quick start

git clone https://huggingface.co/ubanova/UbaNova-16
cd UbaNova-16
pip install -r requirements.txt     # Python 3.9+; CPU-only torch: pip install torch --index-url https://download.pytorch.org/whl/cpu

python generate.py "red potion" "iron sword" --n 4 --model model.pt --palette palette.npy --out sprites.png

Full source (training, web UI, evaluation): https://github.com/irhamuba/ubanova-16

CPU and GPU. The device is chosen automatically: an NVIDIA GPU (CUDA) if one is available, otherwise the CPU. Force one with --device cpu or --device cuda. On GPU the model runs in FP16 autocast; on CPU in FP32.

Tested with Python 3.12, torch 2.15.0, transformers 5.22.2, huggingface_hub 2.4.0, numpy 2.5.4 and pillow 13.0.0 (CPU), and on an NVIDIA T4 (Kaggle).

Useful options:

Option	Default	Meaning
`--n`	4	sprites per prompt
`--cfg`	3.0	classifier-free guidance; how closely to follow the prompt
`--temperature`	1.0	sampling temperature
`--top-p`	0.9	nucleus sampling; removes stray off-palette specks (1.0 disables it)
`--seed`	none	fixed seed for reproducible output
`--scale`	8	nearest-neighbor upscale of the saved PNG
`--device`	auto	`cuda` or `cpu`

The defaults are the best setting from our evaluation. Lower temperatures make shading flat; higher guidance crops objects at the canvas border.

Python

from infer import SpriteGenerator

gen = SpriteGenerator(model_path="model.pt", palette_path="palette.npy")   # cuda if available
images = gen.generate("golden crown with rubies", n=4, seed=42)   # list of 16x16 RGBA PIL images
gen.create_grid(images, scale=8).save("crowns.png")

A local web UI is included in the GitHub repository.

Speed: about 0.05 s per sprite on an NVIDIA T4 in batches (key-value caching), about 1.7 s for a request of 4 sprites, and roughly 6 s per sprite on a 4-thread laptop CPU.

How it works

A caption is encoded by the frozen text encoder of openai/clip-vit-base-patch32 (32 token states).
The model sees [text states | BOS | pixel 1 … pixel 256] and predicts each pixel from the ones before it, in raster order.
Each pixel is one of 4,097 classes: 4,096 palette colors plus transparent.
Classifier-free guidance (10% caption dropout during training) and nucleus sampling are used at inference.


Architecture	decoder-only transformer, pre-LayerNorm
Width / layers / heads	512 / 12 / 8
Trainable parameters	42.44M (CLIP text encoder frozen, not counted)
Sequence	32 text + BOS + 256 pixels = 289
Training data	Angais/pixel-art-atlas-1.4m, 1.41M sprites
Training	3 epochs, 32,934 steps, batch 128, AdamW, lr 3e-4 cosine, FP16, one NVIDIA T4 (~8 h)
Weights	FP16, 85 MB

Evaluation

On 256 held-out captions, scored with CLIP ViT-B/32 (sprites rendered on gray, upscaled to 224 px):

Sampler	CLIP score	Caption retrieval R@1	Colors / sprite	Stray pixels / 1000
Real sprites (reference)	25.68	24.2%	23.6	0.57
guidance 3, temperature 1.0	25.02	16.4%	33.6	0.70
guidance 3, temperature 1.0, top-p 0.9 (default)	25.74	21.5%	22.8	0.50

Held-out loss 1.091 nats/pixel vs 1.059 on training data, so the model does not overfit.
None of 256 generated sprites copies a training sprite; generated sprites are at least as far from the training set as real held-out sprites.
Color words are followed well: red, orange, green, cyan, blue and purple each reach 87–97% on "{color} {object}" prompts.

Rows: red, orange, yellow, green, cyan, blue, purple, pink. Columns: potion, sword, slime, gem, mushroom, shield, dragon, flower.

Known limitations:

Relations between objects are often ignored ("a cat wearing a wizard hat" gives a cat with a crown).
Thin diagonal objects and effects (whips, fire) can come out messy.
Output is 16×16 only.

Free-form prompts, two samples each: a red car; a wooden house with a red roof; a cat wearing a wizard hat; a blue dragon breathing fire; a golden crown with three rubies; a skeleton holding a sword; a cup of coffee; a smiling sun; a treasure chest full of gold coins; an orange basketball; a green frog with a crown; a pixel heart.

Files

File	What
`model.pt`	weights, FP16 state dict (`{"model": ..., "step": 32934}`); loads with `torch.load(..., weights_only=True)`
`palette.npy`	(4097, 3) uint8 RGB palette; index 0 is transparent
`infer.py`, `generate.py`	model definition, sampler and CLI

Data and license notes

Code and weights: Apache License 2.0 (see LICENSE and NOTICE).
Training data: Pixel Art Atlas is released under CC0-1.0, but its images were generated with Krea 2. Krea's Community License says users own their outputs and does not forbid training on them; commercial use of Krea outputs is allowed for organizations with under US$1M annual revenue. Check the Krea 2 license yourself before selling generated sprites (for example as NFTs).
Some platforms require you to disclose that images are AI-generated.
Do not use the model to imitate characters or brands you do not own.

Attribution

If you use UbaNova-16 or build on it, please credit:

UbaNova-16 by Muhammad Fahril Irkham (@UbaNova) — https://github.com/irhamuba · https://huggingface.co/ubanova

and keep the NOTICE file when redistributing.

@misc{irkham2026ubanova,
  author       = {Muhammad Fahril Irkham},
  title        = {UbaNova: A Compact Autoregressive Transformer for Text-to-Sprite Generation at 16x16},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.24214385},
  url          = {https://doi.org/10.5281/zenodo.24214385},
  note         = {Code: \url{https://github.com/irhamuba/ubanova-16}, weights: \url{https://huggingface.co/ubanova/UbaNova-16}}
}

Support the project

If UbaNova is useful to you, donations help pay for GPU time for the next models (32×32 and 64×64).

Network	Address
EVM (Ethereum, BNB Chain, Polygon, Base, Arbitrum, …)	`0x3e888de8c66d2efc9456d730c566563e04ed63e0`
Solana	`9kSnpiHjrrFMrg8PQXAytDYwtjDcDNzc1xDnDq3AxkWY`
Tron (TRX / TRC-20)	`TLvdtLrzknmP4Xjkc59w8UwzGERBDS5ByW`
Bitcoin (Taproot)	`bc1p7hhnfzamcp58n093jfd5ncxz55mejruscwjsclqe5esd86ptxx9q35mqdm`
Sui	`0xa5c680ee5ef209134979eeac30cc192fb5e3cf40cb40483fbca50c229b07d085`

Please double-check the network before sending; funds sent on the wrong network cannot be recovered.

ubanova
/

UbaNova-16