Hugging Face Daily Papers·· 2026-10-19AI score45
Halcyon: a text-to-image generation model without language pretraining
Text-to-Image Generative Modeling Without Language Pretraining
AI briefing
The research team proposed the Halcyon series of early-fusion multimodal autoregressive models, pretrained entirely from scratch on the 100-million image-text pair dataset GPIC, demonstrating that strong T2I generation does not depend on internet-scale language pretraining.
Open original arxiv.org