Scale-RAE

Scales the RAE recipe to text-to-image diffusion transformers from Saining Xie's group, pairing Qwen 1.5B/7B language backbones with 2.4B/9.8B DiTs on top of frozen representation encoders. Weights and the scale-rae-data corpus (~10K downloads) are released on HuggingFace. The models are built on open backbones rather than pretrained from scratch.