ModelsLab/OC LoRA Tests - Isabellav4.0-Augmented

Probably not longer a 'true' single base image anymore. Test version since I didn't spend too much time choosing the controlnet variations and is not as widespread. Only augmented an upper body shot.Uses ControlNet Canny and Lineart to add style augmentation to reduce style bleeding.(helps significantly reduce style bleed and helped with character similarity)this version's approach is somewhat cherrypicked for similar colorsMoved regularization images over to the main dataset using a similar approach found here(helped improve increase character similarity but added more style bleeding)Repeats increase to 10Regularization added as 1_styleN tokens from 2 -> 4Added trigger words for outfit and stylingblock weight removedswapped over to cosine with restartsImpacts:Requires lower weight for general usagebetter character fidelity at lower weights nowhand pose is overfit but is still moveableupper body is somewhat overfitodd artifacts occasionally addedquality of the original outfit is weakerhair color is darker sometimes