Halcyon
#13 in Text-to-ImageStanford · since Preprint arXiv:2610.26863, veröffentlicht um den 20. Oktober 2026 (in News erwähnt) · 2× · last seen Oct 21, 2026
19
Momentum
Halcyon is a family of early-fusion multimodal autoregressive text-to-image models from Stanford and Radical Numerics. The models were trained entirely from scratch on GPIC, a dataset of 100M image-text pairs, without language pretraining. According to the paper, they also learn image captioning (I2T).
Momentum trend
23.07.21.10.
Features
| Fine-Tuning | Qwen3 architecture, random init, trained from scratch |
| Generation Time | Not documented |
| License | Upstream repositories declare no license according to a third-party ONNX repo |
| Max Resolution | 256x256 pixels (GPIC training resolution) |
| Platform | Code and models at cs.stanford.edu/~keshik/halcyon; models on Hugging Face (radicalnumerics) |
| Release Date | Preprint arXiv:2610.26863 (October 2026) |