This model was trained with high quality illustrations by Joseph Christian Leyendecker, with an emphasis on his more "cartoonish" ones. A full list of works used in training, as well as captions, can be found in the training data.
Update 2024.11.02: Reuploaded "Smooth" LoRA, which removed incorrect embedded hashes.
Training Data
The training set was distilled down to ~110 images. Images were cropped and artist signatures were removed. Any horizontal lines, circles or text were manually removed when possible without damaging the composition. Any unfinished studies were edited to remove any distracting secondary elements that may have harmed training. Only a few "Arrow Collar Man" and "Interwoven Socks" ads were part of the dataset due to a lack of image quality and resolution. The full list of works used in training can be found in the training data zip file. Training images themselves can be shared on request.
Description
Trained with SDXL with max total resolution of "1536,1536" in kohya_ss.
Captions in paragraph format were initially generated with Florence-2-Large, however, due to the frequent mistakes, the captions were manually rewritten and expanded to include fuller descriptions and added compositional information (half body, from behind, white background, etc.).
Renders done mostly with JuggernautXL X or AnimagineXL 3.1.
Trigger Word: leyendecker
Many of his works lack backgrounds and so the captions to those images included white background and simple background. A few unfinished studies were included, labeled with unfinished study with guide lines.
For a more stylized image with JuggernautXL X, I found that detailed cartoon illustration by Leyendecker seems to work.
For use with AnimagineXL 3.1, while the details match the Leyendecker illustration style I was going for, the faces may appear too anime-like. If you want more realistic faces, try adding realistic or nose to the positive prompt.