Support me on Ko-fi:
I just released version 2 of this model, you can view the particulars on the right under the "about this version" tab. The following is for version 1:
Description
Dataset trained on 84 images which was a combination of pixel art, official art, and fanart. Unfortunately there are some inconsistencies, mainly in her sash and her thigh highs which may require some reprompting. Also, due to there maybe only being like 3 or 4 images showing her from the rear view, you may get some inconsistent results when trying to generate her from behind. The side also has some issues sometimes but it's generally a lot better. If you're patient enough, however, the model is very flexible and you can do pretty much about whatever you want. Supports both anime style images and pixel art. I only included her base outfit in the showcase, so I'll leave any other outfits (or none :wink:) up to your imagination.
Finetuning (IMPORTANT)
Make sure you have the underscores included in the tags or it won't work properly!
I'm probably going to look into getting a different auto tagger or just make a script to remove the underscores in future, since I don't think it actually makes a difference for kohya ss.
Suggested positives:
tank_top, biker_shorts, (orange_sash:0.5) - (orange_sash:0.8)
Lowering the weight for the sash to between 0.5 and 0.8 appeared to generate more consistent results. You can experiment with increasing/decreasing the weight for the tank top but it didn't appear to make much of a difference. The tank_top and biker_shorts tags are just for reinforcement, not necessary. If you don't want the headwear as pronounced you can either decrease the weight or remove the tag entirely.
Suggested negatives:
buttons, seams, pockets, pink lips, puffy lips
For whatever reason her lips were being generated very exaggerated, it may depend on the model but I had to include those tags in the negatives on every single generation to get it to look right. It also likes to generate random seams/pockets in the clothes, you can experiment with different weights if desired.
Removing headwear:
Positives:
no headwear, no headgear
Negatives:
(winged_headwear:1.5), (headwear:1.5), headgear, headdress, hat, headwear, hair ornament
Removing sash:
Positives:
no sash
Negatives:
orange_sash, sash
The headwear and sash are VERY stubborn, you need extensive prompting both in the positive and negatives to remove them, and even then it doesn't remove it all the time (but it does most of the time).
Conclusion
I'll be honest - this one was painful to make, I had to retrain maybe 5 or 6 times making modifications to the dataset or the tags because it wasn't capturing the features properly, probably skill issue but still. I may go back and try to improve it at some point but it seems unlikely tbh.
Please feel free to share any works you make with this model, either sfw or nsfw, I'd be happy to see what you guys come up with!
Description
Went back and removed some low quality images as well as upscaled some of the existing ones. Should be much more consistent now, but the style has changed slightly, so feel free to stick with v1 if you prefer that better.
I'm not entirely sure why but it doesn't seem to recognize the headgear as well, putting a weird white spot in the middle of it. I've found that decreasing the weight for the headgear anywhere from 0.5-0.8 can help.
Also for whatever reason the pixel art on this version seems a lot more unstable, so I would probably stick with v1 for pixel art or use inpainting with v2. Also changed some of the tags to help the LoRA learn the concepts better.
You can now just use the tags, "no headgear" and "no sash" to get rid of the headgear and sash without extensive prompting, though it may take a few gens.
The prompt accuracy may be slightly worse, but honestly I don't think there's that much of a difference. It should be more or less just a better version of v1.