Halcyon Enables Text-to-Image Generation Without Language Pretraining

Halcyon is a family of early-fusion multimodal autoregressive models pretrained solely on 100M image-text pairs. Halcyon-1.6B outperforms JiT-T2I-XXL/8++ with 275 times less pretraining compute, demonstrating that large-scale text-only pretraining is not strictly necessary for strong T2I generation.