Model Training ‐ Comparison ‐ [Noise Offset]
Models | Logs | Graphs | Configs
Noise Offset (NO) allows you to add additional noise to the training images, theoretically making the generated images more contrasted and vivid.
Compared values:
-
0.00-BD, -
0.05, -
0.10, -
0.15.
DLR(step)
It's indeed strange that the DLR slightly decreases with increasing NO for GR = 1.02, while for GR = ∞, it skyrockets.
Loss(epoch)
And the loss increases in all the cases except for 0.05, which is often recommended to use. However, for GR = ∞ and NO = 0.15, something very creepy seems to happen.
Finally, we have a clear example of an overtrained model with noticeable artifacts :)
Actually, the image periodically becomes more vivid and contrasted. However, in the case of GR = 1.02 with NO = 0.10 and NO = 0.15, the similarity with the character seriously deteriorates, and the colors can become significantly washed out. But now we finally understand what the GR restriction protects us from :) In the case of GR = ∞, good results are achieved with both NO = 0.05 and NO = 0.10, but in the latter case, the initial epochs are terrible, and something of quality is obtained only towards the end of training.
I would say it's safer to use the default value of 0.00, but for the sake of experimentation, you can try 0.05. It's also worth mentioning that SDXL model has been trained with 0.0357 and it could be good choice too. Beyond that, various issues start to arise.
