if you can do better please do this is my first model
Description
Comments (4)
If this is your first time, the best advise I can give you is to either diversify your data or tag for the negative. Adding some 2d art into the dataset can help with letting the lora know how to render the characters in different styles, and so can adding medium tags. There is a lot of Vr chat and 3d modeling software ones in there, so tagging for those will help it learn when to include or exclude the 3d style. Same with the common 3d ones with grid background and multiple views (though there might be a better tag out there for this specific use case of that).
There is some benefit to tagging for other details too, like clothing, expression, background, breast size, angles, lighting, etc, but brute forcing like how you did it is also valid. IL is a pretty flexible model, so you do not have to do that if you do not want to.
Otherwise, follow Danbooru tags for the tagging. You did good for the most part. Just add a dash in "two-tone fur" and make sure when you are using a real tag. "pink hair tips" is not a real tag in danbooru and "colored tips" is more often a tag used for hair that changes color suddenly. I'd use "gradient hair, white hair" and maybe throw "pink hair" in there depending on how I feel.
Also "furry" is a tag you might want to throw in there.
Articles exist on this site for anything else you may need help with.
When you are done, be sure to make test prompts to test the lora under difference conditions.
im more of a visual learner do you know of any youtube videos that could help also thanks for the tips ill look into it when i can. I haven't even used it myself yet since it says download or bid
also how do i take this one and make it better its not in training any more
@skaven I think you can make the lora usable by setting the permission to let people use it in the settings.
As for improving it: Just take your training data and run it through the training process like you did the first time. Be sure to apply any changes you need to make onto the tags or setting first.
You actually made sense of what you were supposed to do the first time around, so I wouldn't even say looking at a video is needed. Although, if it helps, I liked this guy's video on training: https://www.youtube.com/watch?v=C5zpvTG_UZ4
His video may be a bit overwhelming though, and he uses an auto tagger which you honestly do not need if you feel like you can tag the images manually. He places some emphasis on keeping track of what tags you use for consistency. In your case, that would be the copy and paste you already have for every image. You just need to add some other tags to call out important variations and think about how the variety you feed the lora will impact what it will be able to reproduce. All of my generations are coming out looking like VrChat because that's all what you got in there. If you put some normal art of foxyfursure in the folder and tag the 3d ones as "3d, low poly" (which may in hindsight be better than the medium tags I mentioned before), you'd be able to get her in more diverse styles with better lora compatibility.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.