7/1/2026 Anima Update

Anima Update! It's been so long since I've trained anything... But I see this Kanna lora as a special thing for me, so I had to come back at least for this. It has the same dataset, which was mostly due to me being lazy, since it uses E621-style tags. Oops. Either way, it seems to still work pretty good.

Notes:

- For some reason, it seems to focus the image a bit closer to her if used at full strength... Like, images are more likely to be upper-body or headshot portrait without specifying. I have no idea why it does this, you can help fight it by slightly lowering the strength or prompting for 'full length' or 'full body' shots. I'll try and fix this in future versions.
- Hopefully this isn't too overtrained, Anima seems to have very minor knowledge of this character.

- Her 'fishnets' are still pretty inconsistent... Try writing sentences out describing them for possibly better results. Full strength is recommended for any image that has those.

- I have barely used Anima, so hopefully my example images aren't weirdly prompted

- Nova Furry Anima exists. Does it work with this LoRA? Yes! However, my minimal testing shows that it has slightly worse fishnets, at least sometimes. Give it a try!

Older Descriptions:

1/2/2025 NoobAI Update

Happy New Year! I finally got around to retraining this model, and I hope you guys enjoy! This model has a completely new dataset, of which is FAR better than the previous one, this dataset has 373 images, and MUCH better tagging since they are ripped straight from e621. NoobAI already has some Kanna in it's training, but not nearly enough to perfectly replicate her, or her outfit much at all... So this should fix that!

-I trained this new dataset to have considerably better fishnets (netting). It's still pretty inconsistent, but it's definitely a step forward.

-Important: I trained a special tag into the dataset named "kanna classic outfit" for when an image had her standard outfit. Use this tag, and it should improve fishnets in your generations.

-I also trained her "new" outfit she has in the latest game, however so few images have actually depicted in it, that it doesn't really work... Oh well, it was worth a shot. The tag was "kanna alternate outfit" if anyone wants to try it out.

-Also, since Kanna is already somewhat in the dataset, it might not be a bad idea to lower the weight on this, especially if you're mixing it with another lora, or if you simply aren't trying to generate her fishnets (lol)

Older Descriptions:

3/22/2024 PDXL Update:

I updated this model for use with PDXL. Please see the "About this version" section to the right for details. It's pretty much the same model, just trained on better tags, at a better resolution, and with a better optimizer. It's gotten a lot better at doing her face, eyes, and fishnet clothing. Finally, a good Kanna lora! Hope you enjoy.

Older Description:

A lora trained on 235 hand-picked images of Kanna from Blaster Master. It was AI tagged with e621 tags, however it seems to work much better using anime-based .ckpts like Anything 4.5 ...Still though, most models work pretty well with it. If using anime model as the base, try to avoid "looking" tags (looking at viewer, looking away, etc.), or else Kanna might appear to have more anime-styled eyes, and sometimes a nose (see example image #3). Using a full 1.0 weight value seems to do fine.

Some tags like "skintight" and "fishnets" might help achieve her usual outfit. It also handles nudity fine, but I did not include any example images.

I've included a .txt file including all the tags, it's in the "Training Data" download.

4/26/2023 UPDATE: (Now Outdated)

I've trained a new lora on the same dataset, but on a different model than the first version: Fluffyrock.

Fluffyrock is a new "furry" image generation model that has caught my eye by it's surprising quality. It's a lot harder to work with than most anime models in my opinion, but I took the opportunity to train this lora as a test on the new model.

It was trained at 896x896 resolution, again though, using the exact same dataset as the original. It was trained on the fluffyrock 576-832 multires model epoch 20, however, it works well with any of the models, even the new 1088 resolution models.

Fluffyrock downloads: Here (All models) & Here (576-832 e20 recommened)

Notes:
-With the fluffyrock-trained lora, it's much higher resolution, meaning Kanna's eyes are a lot better at understanding expression

-Fluffyrock can handle artist tags quite well, so give them a shot!

-This lora does NOT work with anime-based models at all! Only use this with fluffyrock, use with other models will simply give abysmal results.