OK, I dropped your model into this site:

https://lorainfo.tools/

And the 11 results at the top, some of which have unusually high numbers given the context, are:

"sex": 29,

"1girl": 26,

"hetero": 27,

"penis": 95,

"dog penis": 33,

"glasses": 2,

"vaginal": 12,

"uncensored": 73,

"animal penis": 47,

"outdoors": 21,

"pussy": 12,

and there's 16 "breasts" and 4 "large breasts"

Given that there's 97 "canine penis, I'm gonna take a safe guess that marks the approximate size of your dataset (number of images), So, given that more than a quarter of those are tagged with "1girl" and a seventh or so with "breasts"..

Just sayin that's not a good start. You don't necessarily have to go over your entire dataset(text/tags) with a fine tooth comb, but you can exclude words that have literally nothing to do with what you're trying to train in, if you're using Kohya/WD14 for tagging. Or there's certain text programs (at least in Linux I know) like Kate that will allow you to search for specific words in all the text files in a folder, and replace them. So, you could search "breasts" and replace with a space.

Maybe you're into that kinda thing. I'm not bashing for that, but it's not good for training in this use case scenario. For one, the SD models (SD/SDXL) are SO heavily trained on female images, that you actually have to heavily prompt against quite a number of female attributes(and for male attributes) if you want to consistently generate male content that isn't blended with female parts.

So, you really don't need to train that in, if you just open comfyUI, load almost any model, type "beautiful" and nothing else into the prompt, it will probably generate a scantily clad female (more likely than not). One time I actually had to remove that word from a prompt because I was trying to generate a scene with NO people in it, but it wouldn't stop placing women in my scenes..

That said, I would also suggest going over your dataset (images), and blacking out any breasts, vagina, or anything that isn't directly relevant to "canine penis". If you're creating a character Lora, it's a little different, you can leave backgrounds in and that can add a little something to the Lora, but for something as specific as a body part, especially when you want it to be trained to work on humanoid furry characters, it's best to Just train that part (and the sheath, obviously).

So, to summarize, you want to work on targeting your overall dataset towards the outcome you want to produce.

Oh, and I just saw "antlers", lol.. just 1, but you really want to remove just any words that will do nothing but confuse SDXL(Pony) into generating what you're not trying to generate. That said, there's a lot of "tongue" as well, you'll want to remove all of that, cuz that's why dick loras, especially animal dick loras, often produce dicktongues instead of proper parts.

It can be a pain to go over a dataset of 100+ images, but if you're really looking to create a Lora that shines, that's how to do it. And as I pointed out, there are methods to automate the process for the text.

Final note, just did a quick scroll down over the preview images, it looks like the model is workin alright for people, but one Caveat, I'm noticing some are getting balls and sheath, or balls that have canine penis coming out but lack the sheath appearance. I'd recommend removing "testicles" from the dataset and replacing with "sheath". In theory, that should offer better results.