A LORA for the character Kagura from the Anime/Manga Gintama.

Trained primarily with the goal of getting her "Zunboran Jockstraps" to actually be consistent. Around 40 images, 75 steps/repeats per image or whatever. 100 just deep fried everything, even though most of the guides I read say 100 minimum. Nothing really NSFW in the training. Unfortunately a lot of the (art) training data I had access to were "sketches" which can look good, but they made the LORA impart that "line art"/brushed look on everything, I tried to keep it to a minimum, but what was left mostly had this bright/oversaturated/glow look to it, so expect that so show in the product.

IMPORTANT: DO NOT USE "kagura \(gintama\)" in your prompt with this (or in general really).

Its really bad. The majority of DB images with this tag are either her with the rest of the gang, or some shipping, not many solo. That tag will likely give you a dude with orange "generic male anime hair" and similar clothes. Most of the training had to be cropped from multi subjects.

This LORA was trained on abyssorangemix3AOM3_aom3a3. I was also testing NAI, and anythingV3, but it seemed mostly the same. I use AOM3 mostly, so that's what I went with. Setting it at 1 strength seems fine in my tests.

Most of the training images are with her wearing her cheongsam/qipao or (red) "china dress" in DB tags. Other good tags to use, double bun, bun cover, and side slit (depending on outfit). DPM++2M K is my go to, but I also like S and SDE.

Getting her outfit with the boots is quite hard as DB models all want to put thigh highs or stockings instead of black rubber boots. A challenge for another time.

I HIGHLY recommend you get the "bad-hands-5", "bad_prompt_version2", and "bad-artist-anime" for your negatives if you don't have them already.

I might upload the HN I made that I was using as well, but honestly this does a pretty good job by itself. I might update this, but this was the first Iteration that was consistently good for me. Also it responds REALLY well to highresfix/upscaling where normal 512x512-768x768 are really plain.

EDIT: removed NSFW tag, really wasn't needed