Since I was (originally) not seeing such a model for Illustrious, I present my attempt at a Illustrious character LoRA for Nathan Drake from Uncharted.
V5 is my attempt to further adjust the face details. I honestly don't know what else to do at this point because the training still can't get it as close as I want so this is likely my last attempt because it's "close enough."
V4 is my attempt to revamp the entire model and fix the face details. I was able to get the outfits to work with fewer tags and the face seems more accurate now even if I can't get it 100%.
The tags for V4 & V5 are as follows:
main tags:
uncharteddrake, 1boy, mature maleoutfit tags (3 variants):
casual blue shirt, wristwatch, belt, brown pants, brown shoeswhite dress shirt, suit jacket, bowtie, black belt, dress pants, black dress shoesalternate dress shirt, black belt, dress pants, black dress shoesoptional positive prompt tags:
dirty shirt, shoulder holster, belt pouchLoRA strength recommendations:
Use LoRA strength 0.7 with about CFG 3-4.
The settings I used for my generation testing were 3-5 CFG, 20 steps, DPM++ 2M SDE - Karras on Forge with an adetailer (specifically face_yolov8s).
V3 is my attempt to further improve the face accuracy and details which meant reverting some of my training settings. The LoRA file size did increase but it's the same size as 99% of the LoRAs I already make (just under 200 MB) so nothing different there. I was at least successful in getting his face even closer to being accurate.
The tags for V3 are as follows:
main tags:
uncharteddrake, 1boy, mature male, very short hair, brown hair, blue eyes, stubble, sparse body hairoutfit tags (3 variants):
casual, dark blue shirt, sleeves rolled up, partially unbuttoned, wristwatch, belt, brown pants, brown shoesformal, white dress shirt, suit jacket, bowtie, dress pants, dress shoeswhite dress shirt, partially unbuttoned, sleeves rolled up, dress pants, dress shoesoptional positive prompt tags:
dirty shirt, shoulder holster, belt pouchThe settings I used for my generation testing were 3-5 CFG, 20 steps, DPM++ 2M SDE - Karras on Forge with an adetailer (specifically face_yolov8s).
V2 is my second attempt and produces better results with a more accurate face. I also made the LoRA smaller in file size.
LoRA strength recommendations:
Use LoRA strength 0.7-0.8 with about CFG 3-5.
The tags for V2 are as follows:
main tags:
uncharteddrake, 1boy, mature male, very short hair, brown hair, blue eyes, stubble, sparse body hairoutfit tags (3 variants):
casual, dark blue shirt, sleeves rolled up, partially unbuttoned, wristwatch, belt, brown pants, brown shoesformal, white dress shirt, suit jacket, bowtie, dress pants, dress shoeswhite dress shirt, partially unbuttoned, sleeves rolled up, dress pants, dress shoesoptional positive prompt tags:
dirty shirt, shoulder holster, belt pouchThe settings I used for my generation testing were 3-5 CFG, 20 steps, DPM++ 2M SDE - Karras on Forge with an adetailer (specifically face_yolov8s).
V1 is my initial attempt and produces rather nice results. The face does seem a bit off but no matter how many times I train or change it up the training images (add/remove/etc) it all seems about the same so I'm just posting this as-is (it's close enough for the most part).
LoRA strength recommendations:
Use LoRA strength 0.8 with about CFG 3-5.
The tags for V1 are as follows:
main tags:
uncharteddrake, 1boy, mature male, very short hair, brown hair, blue eyes, stubble, body hairoutfit tags (3 variants):
casual, dark blue shirt, sleeves rolled up, partially unbuttoned, wristwatch, belt, brown pants, brown shoesformal, white dress shirt, suit jacket, bowtie, dress pants, dress shoeswhite dress shirt, partially unbuttoned, sleeves rolled up, dress pants, dress shoesoptional positive prompt tags:
dirty shirt, shoulder holster, belt pouchThe settings I used for my generation testing were 3-5 CFG, 20 steps, DPM++ 2M SDE - Karras on Forge with an adetailer (specifically face_yolov8s).
Description
V5 is my attempt to further adjust the face details. I honestly don't know what else to do at this point because the training still can't get it as close as I want so this is likely my last attempt because it's "close enough."