Trigger word: kyle_hyde
Character model of Kyle Hyde from Hotel Dusk and Last Window.
PONY version:
Activation tag: kyle_hyde
Game 1 outfit: formal black suit, white collared shirt, black necktie
Game 2 outfit: blue jacket, white collared shirt, black necktie
Appearance tags (optional): brown hair, stubble, gray eyes, 34 yo
Trained on CivitAI with default settings (418 steps). Dataset included just 11 images and simplified tags.
After some initial tests:
The pencil drawing style seems to linger, so you may get better results by lowering the LoRa's strength and putting "pencil drawing" in the Negative prompt. If so, you can also use the appearance tags to compensate for the lowered strength. These tags were not included in the dataset so as to make them inherent to the LoRa, but feel free to use them.
---------------------------------------------
SD 1.5 and SDXL versions:
After some testing, both SD 1.5 versions seem to be working fairly well. Mature Male Mix gives some good results from what you can see in the showcase images, but other popular checkpoints also work well.
You might need to include tags such as "masterpiece, best quality etc." in your prompt especially for version E16.
Version E16 (SD 1.5): I used 16 images X 10 repeats X 16 epochs / 2 batch sizes = 1280 steps.
The dataset used for training this model included some sketches / pencil drawings and a few images of lower resolution.
Version E20 (SD 1.5): I used 10 images X 10 repeats X 20 epochs / 2 batch sizes = 1000 steps
In this version I used less images of a better quality, and less black-and-white drawings.
SDXL 1.0 version: trained directly on civitai, with a dataset containing 20 images and updated tags: 20 images X 10 repeats X 10 epochs / 2 batch sizes = 1000 steps.
I need to do more testing but maybe because I included more images (including some of lower resolution) or because I modified the tags, the model seems less reliable in producing high quality results. Plus instead of realistic images it still keeps the pencil drawing style most of the times. It does seem to work better with some checkpoints though - check out the showcase images. Let me know if you can manage to make it work well and consistent.
SDXL E6 version: trained with Tensor Art's online training, with Animagine XL. Updated tags, used only 18 upscaled images of higher resolution: 18 images X 20 repeats, X 6th epoch = 2.160 steps.
SDXL E10 version: same as previous, only it's the 10th epoch (3.600 steps).
SDXL E6 and SDXL10 might be a little under-baked as the character doesn't really look that much as Kyle Hyde sometimes, but it might be better for realistic images. From initial tests, they don't seem to create good results for anime images (they seem to be in a "blurry pencil" style like the one in the showcase).
Civitai link: https://civitai.com/models/277364