UbaNova: A Compact Autoregressive Transformer for Text-to-Sprite Generation at 16x16
We introduce UbaNova, a 42.4M-parameter decoder-only Transformer designed to generate 16×16 pixel-art game sprites from English text prompts. Each sprite consists of 256 sequentially generated palette indices, supporting 4,096 colors plus transparency. Generation is performed pixel by pixel, conditioned on the hidden states of a frozen CLIP text encoder. UbaNova was trained from scratch for 3 epochs (32,934 steps) on the 1.41M-sprite Pixel Art Atlas dataset using a single NVIDIA T4 GPU. The model achieves a held-out loss of 1.091 nats per pixel (1.57 bits per pixel), compared to 1.059 on the training set, indicating no significant overfitting. Using the released sampler with guidance 3.0, temperature 1.0, and nucleus sampling (p=0.9), UbaNova achieves the following evaluation results: CLIP score: 25.74 (real sprites: 25.68). Top-1 caption retrieval: 21.5% (real sprites: 24.2%). Average colors per sprite: 22.8 (real sprites: 23.6). Training set duplication: 0 out of 256 generated evaluation samples. These results demonstrate that a relatively compact Transformer can learn to generate diverse, text-conditioned pixel-art sprites with competitive visual-text alignment. Code: https://github.com/irhamuba/ubanova-16 Model: https://huggingface.co/ubanova/UbaNova-16
Authors
- Fahril Irkham (ORCID: https://orcid.org/0009-0005-9854-2493)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24214384
- Primary Topic
- Generative Adversarial Networks and Image Synthesis
- Type
- preprint