Stanford's Halcyon T2I Family Beats Internet-Pretrained Peers at 275x Less Compute, No Language Pretraining

Stanford's Keshigeyan Chandrasegaran, Fei-Fei Li, Jiajun Wu and co-authors with Radical Numerics open Halcyon, a family of early-fusion multimodal autoregressive T2I models trained from scratch on GPIC, a 100M image-text dataset — no language pretraining. Halcyon-0.52B beats LlamaGen-XL/8 at 18.2x less pretraining compute (48.0 vs. 63.3 gFD); Halcyon-1.6B beats JiT-T2I-XXL/8++ at 275x less compute (35.3 vs. 41.0 gFD). Controlled experiments show text-only pretraining does not improve T2I generation.