Update: Added the LOCON version I mentioned - and the REAL V.3 one - the other one is the very first trained here on Civitai.

Differences:

• Locon: seems to give more freedom and versatiliy, Yet may (chances are it will) need ADetailer for images where focus is bigger than just headshot as usual.

• Version 3: may work better on anime, 3d, semi-realistic models... but may surprise you on realistic models too. You may want that if you prefer a more BARA feels. Although Locon may deliver that too, haven't tried. On LoRA you won't need to type body, nor big hands ands stuff, it will usually make them looking like for you (supposing your checkpoint of choice don't perform it by default, while some "men/bara" models (checkpoints) naturaly add such traits and will get easily how this toon's characteristics should be displayed without extra input.)

• Note: The external sample pics above on LOCON used ADetailer and ESRGAN_4x with steps at 19 and fix at 0.33 - ClipSkip at ZERO or -1.

• On sample pics for V.3, I performed a re-uploading so that now the sample are including full generation data so that you can just copy and play. Yet, all of them are a just 1prompt style.

Locon tips:

- Asian-Less-Neg may be usefull to get his head shape and facial features f you use anime models (for his head is not rounded, and other intrinsic features);

- Since Locon gives more freedom, and it has not been super-trained, nor under-  either, 0.8 is a good weight for it if you can't or won't use any extension simialr to ADetailer (I know some people cannot use them).  It shouldn't interfere with the rest of prompt. If it does, then, yes, lower it, but, please, then, use ADetaler or similar extension if avaiable and add activation tag inside of it - or just gen a pic without it, and inpaint masking the whole head, this time with Locon activated, and moving activation tag and facial/genre promtps to the very beginning. 

LoRA tips:
- It will usually work quite well for body and anatomy, but if weighing it too high IT WILL interfere with clothes and background, unless you use strong prompts/emphasis, as (RED RED RED shirt:1.2), for instance, or "(concert gothic hall background:1.3)" - that if your model has such background in its own dataset. Adding LoRAs which add themes for backgrounds increase a lot diversity. On poses he's been trained with some good variety, hence, he shouldn't be always frontal standing stuck. In any case, a simple ControlNet will make it do. 

- In case you don't have access to any resources above, pick an image you already have and want the result to look something similar, upload to img2img and set threshold about 0.3 to 0.4 (max 0.5), so that your model will generate the new pic using fed information adding new elements you want, but without creating a whole different new picture. For lower freedom (threshold) best samplers are Euler A and Euler (mainly for male SFW), for anything else DDIM and DPM. DDIM and DPM work with lower steps and a bit higher CFG. Hence if Euler A set at 24 steps / CFG 7, DDIM at 17 / CFG 8. Although, unlike DDIM, DPM is more balanced and can be used at CFG 6 or 7 too. You only increase its CFG when steps are very low such as 13 (Same is true for all DPMs... like DPM a and their family)

You can check those pics I upped as samples, and it seems using LOWER CFGs work in our favor. I noticed that when setting CFG higher than 8, instead of making the toon face looking sharper it made it worse, so you can give it a go and see what works better for you. 

First version was a LoRA which was "under-trained" hence even with AD, his face mostly renders out blurry, second time I tried a Locon, twas good for general purposes, has more freedom, but maybe too much freedom, not sure. Still, it helped me to "extract" extra dataset to add to his 3rd safetensor - which without EMPHASIS in the prompt may not allow same freedom/variety, but has great fidelity when it comes to his anatomy.

He can be used for general stuff as a normal character, or for NSFW (it has just a very few pics of him nakey).

Maybe he still needs a 4th version, for now I understand better how to pick up only the good images for the dataset. Nonetheless he already performs very well for many different "occasions".

I have trained him with anime tags, hence 3d and semi-realistic models should work best. Still, I tried him on realistic ones, and it worked fine too. Locon may work better with realistic models I guess (maybe its Locon nature which I'm not so versed about)

You can set his weight very low, such as 0.5 to get the base for body and headshape, then add the LoRA tag inside Adetailer (weight 1.0), with his activation tags there too.

His anatomy (dataset) includes huge hands, large male feet and all bara stuff, but it show itself off what it's got rather on LoRA than on Locon without prompting.

You can check more of his pics on my DA page:

https://www.deviantart.com/dirtyrobbie

He (and all of my toons) has his own separated gallery, so that you can have a look at him only. All other toons have been "tensored" too - although only those who aren't toons made on friends of mine. Those won't ever be turned into LoRAs by myself.

Oh right, in his case, to go straight to his gallery (you will have be logged in, for I tagged not only his, but probably ALL pics as mature):

https://www.deviantart.com/dirtyrobbie/gallery/88507086/juggernaut