More repetitions and slower training basically removed the artstyle at this weight. It was trained on NAI and should be compatible with all anime based checkpoints.

When used by itself I recommend using higher weights for the background, around 20 steps and hi-res fix.

Used with style lora it is pretty consistent at 12 steps even without hi-res fix. Check the examples for more details

Please leave a review if you enjoyed this Lora or have problems with it. I'd also be interested in how the lora behaves with other base checkpoints so why don't you give it a try and throw some images down below!

V1.5

This is a pruned version using a small part of the original dataset.

Use a weight around 0.9. The choker needs to be prompted.

This version is a bit more accurate and can work with together with other LoRAs with some fiddling. For nsfw images I heavily recommend using specific LoRA as can be seen in the examples.

This is my attempt at making a colorized Lora of Miyabi from Miyabi na Senpai by Kyockcho, trained for AOM3. To get the hair and eyecolor right and prevent black and white images I used ControlNet v1.1 Nightly Colab (lineart-anime) to colorize the dataset. Unsurprisingly the model tends towards nsfw images unless specified otherwise.

Weight should be somewhere around 0.9. I have included an XYZ plot for 0.5 to 1 with the basic prompt:

(masterpiece),highres, 1girl,<lora:miyabi-08:1> miyabi, cowboy shot

negative prompt:

(worst quality, low quality, bad hands:1.4)

Seed: 1861942144, 1861942145