About These Models

An attempt at recreating artist Microchip's (HeyChip69) art style as an Illustrious LoRA. It came out okay, but part of the training lost some of the grittiness of the original "sketch" art style. Anyway, hope this one works out for someone's uses. There are two versions:

• Authentic (V2): Trained almost solely on Microchip’s works. Style strongly affects changes to females; most hairstyles will have a very similar shape. I strongly recommend including the term multiple views in negative prompt as the source material is often designed with a comic book styling. In NSFW generation, include dark-skinned male in your negative prompt if this is affecting your desired output. The source material is heavily interracial, which is why it defaults to dark-skinned males.

• Divergent (V1): Trained on hand-picked images (5-10% original artist) and AI-generated images that closely reflected the style. Diverges from interracial source material, somewhat preserves style while enabling broader generations.

Why two versions?

TLDR; I'm still figuring this stuff out. If you are interested in the long answer, keep reading. Originally my concern was that the source material would override and force the model into creating only specific types of images that would limit what an end-user could create (i.e. only interracial). In addition to this, the source material was somewhat "crowded". They contain many panels, text, speech bubbles, and perspectives.

With all of the previous considerations, I figured my best bet was to create and hand-pick images that resembled the style so that the dataset would be more solid. As a result, images lost the tiny details and nuances of Microchip's art style. I felt somewhat unhappy with the final result today, so I continued doctoring Microchip's images until the style itself was preserved but also contained better composition for a trainer to understand.