This LoRA was an attempt to nail down realistic painting, but that attempt failed, so I included some non-realistic figurative artists as well, and got this cool LoRA.
For training I used works of following artists in equal proportions:
- Michael Ancher (Realism)
- Gerda Wegener (Art Nouveau, Fashion illustrations, Erotic illustrations, have standalone LoRA)
- Carl Newman (Drawing, Sketches, have standalone LoRA)
- Suzanne Valadon (Postimpressionism, have standalone LoRA)
- Robert Henri (American Realism, Ashcan School, have standalone LoRA)
- Delphin Enjolras (Academic painting, specifically of dim boudoirs)
- Hans Thoma (Romanticism, illustration, have standalone LoRA)
- Evelyn De Morgan (Pre-Raphaelite, Symbolism, have standalone LoRA on SeaArt)
- Eugene de Blaas (Academic Classicism)
- Carl Reichert (Realism, Hunting dogs, I suspect that some images are actually from different artist)
- Ferdinand von Řezníček (Illustrations, have standalone LoRA)
- Paul Cesar Helleu (Post-impressionism, have standalone LoRA)
- Santiago Rusiñol (Catalan modernism)
From each I took 12 images, except for Thoma - there's 12 paintings and 12 drawings. No artist of movement are tagged - it's a mix of everything and all distinctions come from native for illustrious/noob style and medium tags, overlapping different artists details into single entity.
LoRA name was pulled out of thin air - if you have better suggestions, leave a comment.
V2
Attention. Version 2 is not better than 1. In some aspects it's worse and may require more prompt tinkering. It's less realistic, have stronger training and in the end each artist contribution is unbalanced.
In version 2 I doubled amount of images for each artist (except Thoma) and tagged them. I also added Julie de Graag's prints to the mix. See "about this model" for more details.
As it would require to generate more than 16 thousands images to test all combinations of artists (2^14) just in positive prompt, I didn't do that. I would appreciate if you post some images with your results, leave a comment or review to share your experience.
Description
This one is naughtier, slightly less into realism (up to full anime mode) and with good prompting - less painterly. It's not better than v1.05, it's just different.
Here's what changed.
Added Julie de Graag's prints (have standalone LoRA, v2 trained on Noob on it's way) , increased total amount of images, added artists tags.
Artists tags to use: "ancher", "wegener", "valadon", "helleu", "newman", "de morgan", "de blaas", "de graag", "reichert", "reznicek", "rusinol", "enjolras", "henri", "thoma".
Try different combinations in positives and negatives, some will be better than others. I like "three de" (de morgan, de graag, de blaas) in positive, or to replace "de blaas" with "newman" or "helleu". "de morgan" brings out anime side of things and gives nice yellow-ish or orange-ish tint. Looks cool.
Testing all possible combinations just in positive prompt would require generating 16384 images which on my hardware would require about two full days (45+ hours) . Have fun and post your favorite combinations.