| title | Text-to-Image Generative Modeling Without Language Pretraining |
|---|---|
| authors | Keshigeyan Chandrasegaran, Yue Zhao, Armin W. Thomas, Ehsan Adeli, Jiajun Wu, Juan Carlos Niebles, Li Fei-Fei, Michael Poli |
| date | 2026-10-19 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2610.26863 |
| affiliation | Stanford University |
| upvotes | 5 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf |
| abstract | Modern text-to-image (T2I) systems use internet-scale language pretraining for text-conditioned image generation. We ask whether language pretraining is necessary for strong T2I generation. We introduce Halcyon, a family of early-fusion multimodal autoregressive models pretrained entirely from scratch on GPIC, a dataset of 100M image-text pairs. Halcyon-0.52B outperforms LlamaGen-XL/8 (48.0 vs. 63.3 gFD) with 18.2times less pretraining compute. Further, Halcyon-1.6B outperforms JiT-T2I-XXL/8++ (35.3 vs. 41.0 gFD) with 275times less pretraining compute. Controlled experiments show that text-only pretraining does not improve T2I generation. Halcyon also learns image captioning by replacing caption dropout with I2T supervision, maintaining competitive T2I generation quality under the same pretraining budget. We show that strong T2I systems can be trained from scratch without internet-scale language pretraining. Code and models are available at https://cs.stanford.edu/~keshik/halcyon |
cite
@misc{chandrasegaran2026text,
title={Text-to-Image Generative Modeling Without Language Pretraining},
author={Keshigeyan Chandrasegaran and Yue Zhao and Armin W. Thomas and Ehsan Adeli and Jiajun Wu and Juan Carlos Niebles and Li Fei-Fei and Michael Poli},
year={2026},
eprint={2610.26863},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2610.26863}
}