Loss is taken twice

[landmann]

Hey guys! I'm struggling to understand why the loss is taken twice in the PTI training with the mask. Any pointers? It's different than the traditional dreambooth file, so I'm not sure what's happening.

Thank you!

lora/lora_diffusion/cli_lora_pti.py

Line 366 in bdd51b0

.mean([1, 2, 3])