Can we use Differential Output Preservation loss for LoRA at the moment

[FurkanGozukara]

I feel like it requires special case @bmaltais

Do we have it at the moment? Thank you so much

I mean this pull request : https://github.com/kohya-ss/sd-scripts/pull/1710

So as usual we set regularization images. But it requires a special toml file set. How can we do it? Can we do it?

[DriveHabits]

Its already there you need to make a dataset.toml file instead of img directory

[FurkanGozukara]

dataset.toml

this kills the aim of bmaltais gui automation. can't we add it?

also how we gonna give that toml file from the gui?

[DriveHabits]

@FurkanGozukara no you see the section right next to image folder, Dataset config file, you add it in there, and in that dataset config file you add in your images and regularzation images

[FurkanGozukara]

@DriveHabits can you share here example file that has training images and reg images please

[DriveHabits]

@FurkanGozukara

[[datasets]] resolution = 512 batch_size = 4 enable_bucket = true

[[datasets.subsets]] num_repeats = 1 image_dir = "/home/recris/sd_train/test/export/samples_default"

[[datasets.subsets]] is_reg = true num_repeats = 1 image_dir = "/home/recris/sd_train/test/export/samples_reg" custom_attributes.diff_output_preservation = true

[DriveHabits]

create a .toml file and put this data in there

[FurkanGozukara]

@DriveHabits ty gonna test

[DriveHabits]

only works for lora @FurkanGozukara

[heinrichI]

It can be used with mask conditioning_data_dir ?