It's my maid character. I've found that people tend to think a character isn't real unless there's a something to link to, so I made a LoRA. It works but it's not necessary like, at all, since the prompt alone is plenty sufficient for Illustrious models (that's the point and I'm not a fan of gimmicks anyway). So if you find that this model has a negative style/quality impact on your generations, use a low strength - literally 0 could be a good start. It might be useful to stabilize her in multi-character environments, but tag spillage will still be a problem, so probably not.
Also now available in a hyperrealistic variant by @n_Arno. I have it on good authority that she thinks that's very funny.
~ PROMPT ~
Default look:
1girl, black hair, swept bangs, folded ponytail, gray eyes, glasses, small breastsDefault outfit:
victorian maid, long dress, black dress, latex dress, white apron, maid headdress, long gloves, black gloves, mary janes, white stockingsBut people can change hairstyles and clothes. I'm also making the hunky version official™:
1boy, male, black hair, swept bangs, folded ponytail, gray eyes, glasses, beard stubbleIn a fun twist the LoRA is actually far more useful for that, because Illustrious won't get the dude right without it.
~ PERSONALITY ~
Will do anything but is likely not happy about your bad ideas. People have had varying interpretations of this in the past and I'm not stopping you. I think she's cute.
~ LORE ~
Yes.
Description
Just like with the Illustrious version, this is not required to do the character, so it's mostly just a style LoRA for the look I use the most.
FAQ
Comments (6)
Nice character!
FWIW, IMO the question of "what exists" is best left to the philosophers.
What I ended up doing for one of my OCs is a basic character sheet (front, side, back views on a landscape 16:9). Planning to follow up with the rest once I find the time.
I find this useful for character consistency with the edit mode of models that offer it (such as Flux.2 or Qwen-Image-Edit).
A character sheet is less pointable-at than a LoRA, though, so both approaches have their advantages.
I think regardless of what other people think, a LoRA (even at strength 0) is just the most practical workaround for getting to the block of text tags - I used to have a text file to copy from but it's not as convenient. Additionally, it serves as a tag to automatically collect what images the character is in.
I suppose this could also have been an embedding but I never figured out how to make those.
@John_KSampler Yeah, it's a valid solution. And obviously, if it works for you, it works for you - didn't mean to imply otherwise. :)
The block of text tags? Don't those still need to be entered manually even with a LoRA? (At least with the builtin LoRALoaderModelOnly node.)
Automatic collection, good point.
Embedding, yeah, that's pretty much the tool that matches the shape of the problem - a compact keyword to elicit what the model already knows. I haven't made those either, so unfortunately can't help there.
@Mathemagic This is mostly for the online generator, but for Comfy I always use rgthree's multi-lora nodes, which hook back to CivitAI to give you the tags.
@John_KSampler Thanks for the pointer!
I've been trying to minimize dependencies, so I haven't tried those.
After all these years, I'm still pasting in from a text file. Or from a previous image. :)
@Mathemagic Have you tried Lora Manager? It scrapes the example images and lora descriptions of a model and you can browse them locally in a seperate tab (It has search and filtering capabilities too for large lora collections). Also supports checkpoints too, i used to copy and paste everything into text files as well and it's been a huge timesaver for me. (I only use the single Lora loader node from it in my workflow but it comes with a set of replacements for most common nodes for some reason.)