Got a request for Louy from Sprunki Toxic (OC-ish). Had only one image to go by so I needed to get a bit creative about it
Trigger: l0uy_st
Helpful: female, crystal_creature, chest gem, dress
Noticable issues:
Low image LoRA means it's volatile, makes it irregular and adds/removes human features while trying to keep the head-style
Helpful suggestions: repaint and img-img/inpaint, running at lower (0.7-0.9) strength or higher (1.1-1.3)
Description
Comments (6)
Try my method to learn how to create a better LoRA of a character with scarce images!
Take the character's already-existing image(s), make multiple cropped versions of it, take screenshots/a screenshot of a 3D Sketchfab model and trace or edit the screenshot to look like the character if it doesn't already depict them, generate AI images with the character if you can (via text prompt, img2img, or an already-existing LoRA that might be low-quality but help for the dataset), and take many images from the internet that can be edited/traced to look like your character. Put them in your dataset to train your LoRA, and it will come out better and more flexible than an LoRA trained on only one image.
I used this method to create an LoRA of my favourite OC who only unfortunately had a few images to use for my dataset, and it benefitted from this way. You can generate him standing or sitting, viewed from the front or back, and in a simple or detailed style. It can also be used for other purposes, such as to enhance quality, and for a certain body type. Cool, right?
Yeah I used Grok and ChatGPT to generate a few 3D views and somewhat humanized forms. When this is all I had to go by it was a rather rough request.. https://i.imgur.com/Dg2VVji.png
@xerberus0For more tips, try making images of them in many instead of just a few views! Angles and nudity/revealing levels of outfits. If you want to draw the character yourself or draw over images from the internet with many different views, it doesn't have to look perfect; AI just needs an idea of what to generate with the LoRA.
However, you also have to balance the quality of your processing with removing anything harmful to AI understanding the character of the LoRA. For example, the Shadow Skitty produced by CitronLegacy's Shadow Skitty LoRA in an experiment looks fluffier than the original-forme Skitty, which is possibly caused by non-precise outlines in the edits. The edits used for the dataset were sloppy and AI might have thought of them as having fuzzier hair.
You might also want to use img2img to generate the character in multiple styles, to enhance flexibility of your LoRA. Alternatively, you can take existing images without the character and edit their characters to look like the desired character, or trace on them with different kinds of shading.
@urbanlegendwiki Yeah, no, I get that and I did "anthro" the character as humanoid as possible for some variety as well as images, however I can only do so much with the sepcific layout of the crystals, especially since it doesn't have any level of fanart aside from eye changes and blushing, which I did include
@xerberus0 If you mean that you worried about how AI didn't get the specific crystal layout right, that's why you shouldn't only rely on AI image generation and gathering already-existing images of the character! Even if you know the character has a scarce amount of manually made already-existing images, you can also take an image from a booru, trace over it, and draw accurate crystals to avoid training the LoRA on inaccurate crystal layouts. You can also manually edit the AI-generated images' depiction of the character's crystal layout to make them more accurate.
@urbanlegendwiki Yeah more or less what I did, aside from tracing stuff I suppose. I'm not being facetious when I say it literally has one image with a few facial changes. Overall pretty satisfied with what it became in the end, considering everything