ai-forever / Kandinsky-2

Kandinsky 2 — multilingual text2image latent diffusion model
Apache License 2.0
2.77k stars 306 forks source link

Batch size for prior training #90

Open eladrich opened 1 year ago

eladrich commented 1 year ago

Hi, Great work with the Kandinsky model, the last improvements look really impressive 🎨

For prior training/tuning I saw that the default batch size is 1, is that actually the size used during training, or is a larger batch needed for stable training? Would it be possible to share the configuration used for training the prior from scratch (the one that took 1M iterations)