Loss almost does not change

[Oguzhanercan]

Hi, I am trying to experience different losses. I Have implement a face similarity loss and disabled all of other losses. But The loss almost does not changed (%1 decreased). At my method, I am looking at the cosine similarity between the reference image and generated image like RectifID(https://github.com/feifeiobama/RectifID). Do you have any experiments with that or do you have any suggestions about it?

Face recognition model: Arcface Resnet200 Loss: (1 - Cosine simlarity) Lr: 1 to 30 (I tried different settings on this) SD Model: Pixart