Paul César Helleu was a French oil painter, pastel artist, drypoint etcher, and designer, best known for his numerous portraits of beautiful society women of the Belle Époque.If you live in New York and happen to be in Grand Central Terminal, check out his celing mural of constellations

inspired by Paul Cesar Helleu

This version replicates drawing style of Paul Cesar Helleu, although final result will heavily depend on you used checkpoint and prompt. For example, usage of colors will diverge from "pencil drawing" towards "painting", but that's not bad, and can give interesting results. Not all V-prediction checkpoints will work, in my testing some are affected by this LoRA, but images do not resemble Helleu's drawings or paintings at all. It will work with some Epsilon prediction checkpoints, but again, not all of them. On Epsilon results may be of entirely different style, but that is not necessarily bad.Experiment, try different things, most importantly, have fun. Let me know what you like and don't like, post some images with your favorite checkpoint.Things to try:<ul><li>Color themes (like blue-theme, pink theme, etc)</li><li><code>monochrome</code>, <code>greyscale</code>, <code>monochrome, greyscale</code></li><li><code>sketch</code></li></ul>Potential problems:<ul><li>Tendency towards female characters, so may genderswap or crossdress your character</li><li>Default color hair comes from Helleu's palette, so it's somewhat red hair with black or grey</li><li>Faces may be awful</li><li>May undress your character</li><li>May give your character a dress</li></ul>Not problems:<ul><li>Signature. With my LoRAs it's a feature</li><li>Image may fade out into white loosing details or leave some parts "unfinished". It's from artist's style</li></ul>Tips:<ul><li>Limit use of color</li><li>Don't make complex environments or backgrounds - there is no such nonsense in Helleu's drawings</li><li>Experiment with prompt detalisation</li><li>Have fun and post an image or leave a comment</li><li>Play with LoRA strength. <code>1</code> is a good basis, but between 0.5 and 1.3 or 1.4 can give better results depending on prompt and checkpoint used</li></ul>Tested with (and had nice results):<ul><li>Noob AI - both V-pred and Eps give great results. Trained on V-pred 1.0</li><li>Pasanctuary 5.0 (4.0 works, but weird and not in style)</li><li>Obsession V-pred 1.0</li><li>WAI-SHUFFLE-Noob - result is nice mix of anime-style and Helleu</li><li>DOMKU v2 - light effect</li><li>Hassaku XL Illustious 1.3a and v2 - may give interesting results, or may give "meh"</li><li>PencilXL V-pred - works, my favorite</li><li>Cat Tower V-pred - doesn't work at all</li></ul>

Files with SHA256: 45c122db9c77e8461e5d2d9078a10c45e88a74b37a365bf19de33e3d6ccf6983