Stanford's Halcyon: text-to-image without language pretraining, beating rivals with up to 275x less compute
serrjoa · x · 2026-10-21
A Stanford team (including Li Fei-Fei) introduces Halcyon, a family of early-fusion multimodal autoregressive models pretrained entirely from scratch on 100M image-text pairs. Key findings: text-only language pretraining gives no benefit for T2I generation, and strong T2I models can be trained without internet-scale language pretraining. Halcyon-0.52B beats LlamaGen-XL/8 (48.0 vs 63.3 gFD) with 18.2x less pretraining compute; Halcyon-1.6B beats JiT-T2I-XXL/8++ (35.3 vs 41.0) with 275x less. Swapping caption dropout for I2T supervision adds image captioning at no extra budget. Code and models are open-sourced.
Related event: Stanford's Halcyon: Text-to-Image Without Language Pretraining(2 posts)→
More from Research
- Laude Institute launches FrontierCS 2: agents must beat deleted paper contributions blind — lateinteraction · 2026-10-23
- RFD4-Proteina's 'underrated' docs detail partial diffusion for protein redesign — AllThingsApx · 2026-10-23
- MPI-Sec PhD Thesis on Vulnerability Detection Benchmarks Includes Pre-LLM Classic "Top Score on the Wrong Exam" — mboehme_ · 2026-10-23
- LiveNerf's 29-day data finds no nerfing of Claude Opus 5.5 — TheOnlyVibemaster · 2026-10-23
- New paper makes 9 falsifiable predictions toward a major AI-enabled terror attack by 2027-2028 — StephenLCasper · 2026-10-23
- HumanoidPose recovers humanoid robots' full-body joint states from RGB video alone — cholz · 2026-10-23