Scaling handbook for text-to-image on 100M pairs: retire ImageNet, drop CFG and EMA

DimitrisPapail · x · 2026-11-17

Wenhao Chai published "How to Scale Text-to-image Models: a Handbook," training pixel-space JiT models from 3.6e18 to 2.3e20 FLOPs on GPIC (100M permissive text-image pairs), bringing LLM pretraining science to visual generation.

Key findings:

All code, configs, and sweep data are open-sourced.

Original post →

More from Multimodal

Multimodal channel →