Adv_loss don't reduce for Flux Implementation

[okus-hubx]

I’m experiencing a problem where the adversarial loss (‎adv_loss) does not decrease during training when using my Flux implementation. I think my d_loss is reducing more faster than expected. Did you come across problem like this ? or Can you give advices to solve this problem ?
(I am using one H100 so my batch size is 2)

[Run-19]

Hello, I have also tried to reproduce the results using FLUX. May I ask if you have successfully reproduced it? I used distilled weights for inference and found that the results were better than those obtained by directly using the teacher model with fewer inference steps. However, there is always some noise that cannot be removed. Have you encountered this situation? Do you have any suggestions for this？

[okus-hubx]

Hi, Actually I continue my experiments on distillation. I also encountered same noise problem and trying to solve it. But i don't like my outputs quality when I comparing it with schnell. so before starting to solve noise problem, I am trying to achieve same quality as schnell. Did you compare your outputs with schnell? Also if it's not problem for you, can you share your training setup.
(dateset size: 100000,
traing step: 100000,
batch size:4,
device:H100,
lr:1e-6)
I'm thinking of increasing my dataset size.