Halcyon: from-scratch early-fusion T2I models match late fusion with 275x less pretraining compute
StanfordAILab · x · 2026-10-16
- Radical Numerics (with Stanford AI Lab researchers) released Halcyon, a family of early-fusion text-to-image base models trained from scratch
- Core question: does T2I need language pretraining? Halcyon shows it can skip the separate language model and still beat strong late-fusion baselines with up to 275x less pretraining compute
- Context: in both late and early fusion designs, language pretraining dominates the compute budget; this work studies how modalities should be combined and compute allocated
- Models, recipes, and paper are open; part of a research program on multimodal pretraining science (incl. Omnii genome models)
Related event: Halcyon: Text-to-Image Base Models That Cut Pre-training Compute 275x(3 posts)→
More from Multimodal
- Open-sourced 61-second computing history film generated entirely by Muse Spark — sophiamyang · 2026-10-16
- Runway launches Replit integration to generate images and video inside web design projects — tlakomy · 2026-10-16
- Atlas demos turn a single image into an explorable 3D scene, generating off-screen areas live — XRarchitect · 2026-10-16
- HyperFrames launches Reference Any Video and auto-applied Brand Kits for AI video — aftahi_ai · 2026-10-16
- Krea 2 + LoRA still beats Flux 3 Klein in pixel-art style test — Angaisb_ · 2026-10-16
- World Labs starts rolling out Atlas beta with pixel-perfect camera path control — theworldlabs · 2026-10-16