Halcyon: Text-to-Image Generative Modeling Without Language Pretraining

Forum topic · 小凯 · 2026-10-21

Summary

A forum post introduces Halcyon, a family of early-fusion multimodal autoregressive models for text-to-image (T2I) generation pretrained entirely from scratch on GPIC, a dataset of 100M image-text pairs, questioning whether internet-scale language pretraining is truly necessary for strong T2I performance. According to the paper (arXiv:2610.26863), Halcyon-0.52B outperforms LlamaGen-XL/8 (48.0 vs. 63.3 gFD) using 18.2x less pretraining compute, while Halcyon-1.6B surpasses JiT-T2I-XXL/8++ (35.3 vs. 41.0 gFD) with 275x less pretraining compute. Controlled experiments indicate that text-only pretraining does not improve T2I generation quality. The authors also show Halcyon can learn image captioning by replacing caption dropout with image-to-text (I2T) supervision while maintaining competitive T2I quality at the same pretraining budget. The work, authored by researchers including Keshigeyan Chandrasegaran, Yue Zhao, and Li Fei-Fei, demonstrates that strong T2I systems can be trained from scratch without internet-scale language pretraining.

Paper Overview

Research area: Computer Vision Authors: Keshigeyan Chandrasegaran, Yue Zhao, Armin W. Thomas, Ehsan Adeli, Jiajun Wu, Juan Carlos Niebles, Li Fei-Fei, Michael Poli Published: 2026-10-19 arXiv: 2610.26863

Abstract (translated)

Modern text-to-image (T2I) systems rely on internet-scale language pretraining for text-conditioned image generation. The authors ask whether strong T2I generation actually requires language pretraining. They introduce Halcyon, a family of early-fusion multimodal autoregressive models pretrained entirely from scratch on GPIC, a dataset of 100M image-text pairs.

Key results:

Halcyon-0.52B outperforms LlamaGen-XL/8 (48.0 vs. 63.3 gFD) with 18.2x less pretraining compute.
Halcyon-1.6B outperforms JiT-T2I-XXL/8++ (35.3 vs. 41.0 gFD) with 275x less pretraining compute.
Controlled experiments show that text-only pretraining does not improve T2I generation quality.
Halcyon also learns image captioning by replacing caption dropout with I2T (image-to-text) supervision, maintaining competitive T2I generation quality under the same pretraining budget.

Conclusion

The paper demonstrates that strong T2I systems can be trained from scratch without internet-scale language pretraining, challenging a common assumption in modern generative modeling pipelines.

--- *Auto-collected on 2026-10-21*