Activation Tag: bbell
Reinforcement Tags (strongly recommend, optional):
anthro, furry female, animal ears, ox tail, yellow horns
pink fur, two-tone fur, white fur, pink hair, short hair
white shirt, pink skirt, white gloves, white shoes, blue baseball cap, no headwear
blue scarf, pink long sleeves, long scarf
baseball bat, baseball mitt
white bloomers, bloomers under skirt
Negative Tags (all optional):
mascot (training dataset uses some photos of the actual mascot, which I have added with the tag exclusively for the negative prompt)
chibi (just a stylistic choice, if you want better anthro proportions)
text, icon, shirt sign (do this if you feel the gibberish text is too much; see below)
Recommended settings (green text is what was used in my personal preview images):
LoRA strength: 0.7 - 1.0 (0.9)
Steps: 20 - 30 (27)
CFG: 4 - 7 (5.0)
Scheduler: Euler A / Normal OR DPM 2M ++ / Karras (Euler A/Normal)
Been awhile since I made my first character LoRA, and this is a culmination of what I learned since. This LoRA is a bit different in that it uses fewer training tags and is trained to heavily rely on the activation tag to put the character in the image. This should allow you to spawn her without having to use so many tokens, and still have the flexibility to change her clothes if need be.
It is tempting to use her name (buffalo bell) as a reinforcement tag, but I HIGHLY RECOMMEND OTHERWISE. Most base/merge models simply do not have good data population of that character, and will result in wildly inconsistent results. Use the reinforcement tags I listed above instead.
The training images were selected for accuracy, but leans towards a very cute anime aesthetic. This is a tradeoff for using a large amount of training steps on a small set, and is unintentional. If this bugs you out, use the anthro tag in your prompt and consider using body tags.
As a fair warning, the LoRA suffers from the text/icons being gibberish on the cap and shirt. This is unfortunately unavoidable at this moment. I might consider manually editing them out of the images if there is enough interest in this LoRA.
For those that know the mascot, know that by default the activation tag gives her the summer outfit. If you want the winter variant, add blue scarf, pink long sleeves, long scarf to your prompt.
Alternatively, you can put mascot in your prompt if you prefer that, just be warned that the outlines may not be consistent and that you may not be able to swap her facial expressions. Use DPM 2M ++ / Karras for better, more consistent outlines.
The LoRA should be sufficiently trained to hold most of Buffalo Bell's features, but in anime-centric models and with too many tokens, she might default to being turned into a generic anime woman. For best results, use a furry-centric model or use the first two sets of reinforcement tags listed above. I will recommend some models below that I personally find exceptional for this.
Description
This version is actually v2.0, but v1.0 uses an old tagging system with erroneous extra activation tags. Please see the page description for best tagging practices.
FAQ
Comments (7)
Always happy to see a Buffalo Bell model
She's so damn cute
I did not know any sports team had a mascot like this. Cute.
Quick advise: Populate the preview with shots of her looking older or you may receive censoring due to decisions far beyond your control.
Oh, I did not know that! I went with what was closer to the mascot's look because I thought it would make the design more consistent.
@aggresive I know right.
You are not at fault. Civitai rolled out a bunch of features to appease payment processors after they momentarily shut down payment to the site. They ended up going way too far though and did stuff like this that punished SFW use of the website.
@aggresive On the topic, your use of the tag "v sign" where the actual tag is "v" gives me a good excuse to talk about how these things are trained on Danbooru tags. I do not know if you knew that already, but this one or any of the breast tags above "small" may come in handy. You can add them to to the tag list as an incentive.
@Diamond_Letter It took me mostly trial-and-error to learn about what not to tag in this model, and I feel I can still improve on it. "closeup" for example gets way too amplified in this model. I'd like to know if you have any article recommendations for tag usages in LoRA training.
@aggresive This one is pretty good: https://civitai.com/articles/8731/my-personal-lora-creationposting-guidelines
The only detail he leaves out is framing tags like portrait, full body, upper body, feet out of frame, cowboy shot, head out of frame, etc.: https://danbooru.donmai.us/wiki_pages/cowboy_shot
Close-up as a tag is mostly intended as an add-on to some of those tags to show it is really zoomed in on the subject. So for example, if you tagged "close-up" for a portrait shot, that would only be needed if her face was taking up the whole frame.
I also have a few program recommendations and tips in my article that can help:
https://civitai.com/articles/26802/tagging-data-prep-methods-and-tips
Apart from that, it's only really a matter of checking to see if you are using the right tags. You seemed to have done a really good job for the large majority of them. I did not see this before, but the only thing that jumps out at me in your listed activation tags is to change "anthro" to "furry". It's a common mistake, since I think that is an e621 tag, which some models use. I think you can remove "bloomers under skirt" too.