Stanford's Halcyon: strong text-to-image without language pretraining
StanfordUniversity · hf · 2026-10-20
A paper from Stanford and Radical Numerics (Li Fei-Fei, Michael Poli, et al.), Halcyon questions the default assumption that strong text-to-image generation requires internet-scale language pretraining.
- Introduces 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 63.3 gFD) with 18.2x less pretraining compute; Halcyon-1.6B beats JiT-T2I-XXL/8++ (35.3 vs 41.0 gFD) with 275x less compute
- Controlled experiments show scaling text-only pretraining doesn't improve T2I; replacing caption dropout with I2T supervision keeps competitive generation under the same budget while learning captioning
- Code and models available at cs.stanford.edu/keshik/halcyon
More from Multimodal
- Qwen releases BlenderWorld benchmark for multimodal 3D creation agents — Qwen · 2026-10-20
- Stunning AI-generated gothic portrait wows the AI art crowd — Loo_Atreides · 2026-10-20
- ComfyUI hits 140k stars: text-to-image generation now runs on your own GPU — alex_verem · 2026-10-20
- SwitchSplat brings switchable object states to 3D Gaussian Splatting, headed to SIGGRAPH Asia 2026 — taiyasaki · 2026-10-20
- FastVideo releases FastH3 OmniRef for ComfyUI on Hugging Face — Lost-Dot-9916 · 2026-10-20
- Qwen 2.1 Shows Worse Subject-Scene Integration Than AIO v23 in Multi-Image Edit Tests — ShowerFluid9178 · 2026-10-20