Style training settings
First off, wow I love this LoRA. I really appreciate the way you've identified problems with the base model (to much dof, "pro" photography stylization) and tackled it with this.
I'm just starting to experiment with Flux LoRA training after training a few SDXL LoRAs, and I wonder if you would be willing to share your config for training style. I'm curious about how many images are in your dataset, learning rate, DIM and ALPHA. Thanks in advance.
These LoRA's I believe were made with the default configuration of Ostris's AI-Toolkit with maybe the batch size increased to 4. It was using the default config from probably a commit two weeks ago on ai-toolkit. Learning rate may have been at 4e-4 at that time time. DIM and ALPHA probably at 16 though you may get better results at 32 and higher.
I think there were probably around 90-120 images in this dataset, but there is a decent chance you can get better results with less images.
I usually like to stick close to default parameters for trainers with reliable results whenever possible, as it is easier to trial and error the different significant aspects of the dataset alone.
Thanks so much, I'll try this. Sticking with the default settings at first is probably a really good tactic.