This adaption was trained to work with Think Diffusion XL.

((high rank & low alpha, no= ~0.018, clip skip:1, BRS=64, Timestepped for 100% noise (& not ideal for Img2Img I am also planning on releasing an Img2Img), SDXL only; optimized for 1024 x 1024, 640 x 1536, 768 x 1344, 832 x 1216 , and 896 x 1152.))

RECOMMENDED SETTINGS

Sampler: Euler A & Steps: 15 to 30. with the sweet spot being 18 to 22.

CFG: This adaption has a lot of dimention to it as a result it should punch above its weight less focus on the model and more on the classifier guidance should be higher than 15. CFG 18 to 21 produced excellent results, some results above 21 were superior to below but it was very speicfic cases in which I would consider it fine tuned and inflexible.

Denoising Strength (img2img only): less than .85 often turns faces into sculptures, (.85) produced more realistic results with many lora strengths,

Lora Strength: No less than 1 is recommended. 1.15 to 1.25 was good.

Things that tested and confirmed work well:

Positive Prompts: "Caricatures style, solo, portrait, simple background, looking at viewer, realistic, shirt, monochrome, greyscale, white background, sketch, upper body, signature, nose, collared shirt, traditional media,jacket, white shirt, thick eyebrows, formal, parody, artist name"

(removing parody or realistic should change the output, moving signature and artist name to negative should hopefully remove signatures they are features of this style I prefer them)

facial features that seem to work good : closed mouth, smile, lips, teeth, black eyes, brown eyes, open mouth, grin

Gendery things: male focus, 1boy, 1girl, female focus, jewelry, manly, female focus

Age: old man, old, wrinkled skin, young, child

Hair: facial hair, mustache, short hair, long hair, beard

"Caricatures style drawing of null, null2, solo, portrait, simple background, looking at viewer, realistic, shirt, monochrome, greyscale, white background, sketch, upper body, signature, nose, collared shirt, traditional media,jacket, white shirt, thick eyebrows, formal, parody" =was used for a lot of testing

Good Negative prompts: photo,Realistic portrayal,Subtle representation,Naturalistic depiction,Authentic portrayal,True-to-life rendering,Unexaggerated portrayal,Factual representation,Literal depiction,Accurate likeness,Abstract art,Impressionism,Realism,Minimalism,Surrealism,Cubism,Pointillism,Hyperrealism,Photorealism,watermark

IMG2IMG model: ((high rank & low alpha, no= ~0.018, clip skip:1, BRS=64, Timestepped for 300min/700max noise, I was going to release this separately but due to losing a large amount of research data this is released at the same time as the original, SDXL only; optimized for 1024 x 1024, 640 x 1536, 768 x 1344, 832 x 1216 , and 896 x 1152.))

Haloing original image as background color is common and is considered a feature of this style, albeit I have not perfected it yet, you can remove the background to prevent this. It should go without saying that with a CFG this high if your model is familiar with the person it will output better, so famous people will have a better img2img then your family photos. Controlnet might improve that but I have not tested it due to an issue with a VAE that rendered several days of research useless.

Common mistake: "Caricature style" instead of "Caricatures style" feel free to use both but you should use the one with S at the end.

Image Selection for img2img: doesn't matter I used stable diffusion generated portraits for one celeb and the same image worked fine for 15 others. The img2img version is a lot like controlnet. Maybe if you get it to work with lower denoise the opposite can be true in testing I did not like it with low denoise.

I would also just like to say this is my first published LORA in a very long time that does not replicate a person or used to create porn. Of course if your checkpoint is trained on those things you can probably come up with some funny stuff with this. Dozens of my original LORAs were removed back in April and I decided it wasn't worth my time to make my subsequent work public but I love to contribute as well. I don't see the distinction with celebrities and for some reason some celeb loras are ok and some are not, whatever I'm happy we have what we have.

anyway this LORA is not trained on any specific person by name it is trained in the art of caricatures and parody which is actually pretty complicated and I can't draw this stuff myself which is why I wanted to make it so everyone could reproduce decent ones you are going to want to save your best seeds with this!

I also have a B an d C version for all that I can release if you want in my experience they were very similar and I didn't want to put out 6 versions of the same lora.