Jmoz style ILLUSTRIOUS

This is a LORA to force ILLUSTRIOUS to generate images in the style of Jmoz. This is a cartoonish style and ILL is extremely biased towards generating anime-styled images, but it works :)

When using this LORA, note the following:

1. Different ILL models have different levels of style "strength" baked into them. The best results of using this LORA in my testing were achieved with ILL_Hassaku. Other models I tested tried to mix the Jmoz style with their "native" style, so what you'll get is a kind of a mix. And they also garble faces pretty badly.

2. Often when generating, you may get bad faces, when the model tries to mix an anime face with the cartoonish face from the LORA. This is resolved by upscaling the image. Upscaling fixes any defects in the image very well, so don't be afraid if you get a mediocre result on the first pass.

3. ILL (based on Danbooru tags as it is) actually knows Jmoz as an artist, but it's native dataset is very small (this LORA uses 100 newer images by Jmoz). So it helps if you tag "jmoz" in the prompt.

4. The LORA is not very hard-baked, so use it at strengths around 1 to model's GUI around 3-4. Stronger GUI of the model will force more native model style, overriding the LORA. Lowering the LORA stregth achieves the same effect. Could be used to "spice-up" the native ILL anime style with a little crazy toonishness. LORA strenghts over 1.3 will typically garble the image, but your mileage may vary.

5. The dataset that I used for training was fully tagged. This means that LORA will handle the things that were originally present on Jmoz's images better and things that he never did draw will be more based on the original model.

P.S. of course don't forget the idiotic quality tags that need to go with every ILL generation, unless you want to generate a bad pic.

I personally use:

Positive: good quality, masterpiece, best quality, 8k, amazing quality, very aesthetic, highres, absurdres, detailed

Negative: worst quality, bad quality, lowres, (bad), bad anatomy, bad hands, extra digits, fewer, extra, missing, text, error, worst quality, jpeg artifacts, low quality, unfinished, displeasing, early,artistic error

(Though I'm pretty sure half of these have no effect, I never had the time to test them out properly. But overall they work.)

That seems to be all :)

Happy burning!