question about balanced l1 loss

mit-han-lab / fastcomposer

[IJCV] FastComposer: Tuning-Free Multi-Subject Image Generation with Localized Attention

https://fastcomposer.mit.edu

MIT License

641 stars 36 forks source link

question about balanced l1 loss #28

Open lwb2099 opened 8 months ago

lwb2099 commented 8 months ago

I have a few qustions about this loss: 1. threshold is no use, why; 2. if two attention maps are the same, loss should be -1, but shouldn't loss be optimized from somthing positive to neat zero?

xilanhua12138 commented 1 month ago

same question