Paper recorded by Signals 4 on 2026-10-19 in cs.LG. Abstract reproduced from arXiv; link to the original below.
Published 2026-10-19 on arXiv · recorded by Signals 4 on 2026-10-20
Category: cs.LG · 机器学习 · first seen 2026-10-20
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