the Training details of ella

[icelighting]

ELLA 后出来的embedding 和diffusion model的 alignment 表现是怎么表现的，可以讲一下吗？谢谢～

[DthdZK]

同问，请问在训练的时候训练的损失函数是如何设计的呢?

[budui]

ELLA only used the Diffusion MSE loss.

Replace the code in https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_sdxl.py that optimizes the UNet with code that optimizes the connector instead.

[DthdZK]

Thanks a lot! I'm going to try to reproduce the training.