1B to 4B parameters from scratch on GPIC, including the text tokenizer, image tokenizer, and generative model, using next-token prediction over interleaved image-text sequences.
Strong text-to-image generation is possible without any language pretraining
Chandrasegaran et al. show that Halcyon, trained from scratch on 100M image-text pairs, outperforms larger systems that rely on pretrained language encoders while using up to 275 times less pretraining compute.
Top University
Keshigeyan Chandrasegaran · Yue Zhao · Armin W. Thomas · Ehsan Adeli · Jiajun Wu · Juan Carlos Niebles · +2 more
Stanford University · Radical Numerics
Research Digest··2 min read
Chandrasegaran et al.
§