mirror of
https://github.com/ostris/ai-toolkit.git
synced 2026-04-28 10:11:14 +00:00
Added Differential Output Preservation Loss to trainer and ui
This commit is contained in:
@@ -341,6 +341,12 @@ class TrainConfig:
|
||||
# unmasked reign. It is unmasked regularization basically
|
||||
self.inverted_mask_prior = kwargs.get('inverted_mask_prior', False)
|
||||
self.inverted_mask_prior_multiplier = kwargs.get('inverted_mask_prior_multiplier', 0.5)
|
||||
|
||||
# DOP will will run the same image and prompt through the network without the trigger word blank and use it as a target
|
||||
self.diff_output_preservation = kwargs.get('diff_output_preservation', False)
|
||||
self.diff_output_preservation_multiplier = kwargs.get('diff_output_preservation_multiplier', 1.0)
|
||||
# If the trigger word is in the prompt, we will use this class name to replace it eg. "sks woman" -> "woman"
|
||||
self.diff_output_preservation_class = kwargs.get('diff_output_preservation_class', '')
|
||||
|
||||
# legacy
|
||||
if match_adapter_assist and self.match_adapter_chance == 0.0:
|
||||
|
||||
Reference in New Issue
Block a user