Male Mix Illustrious Edition

Finally, it is here. My Male Mix Lora has moved to Illustrious.

Originally I didn't have plans to rebuild this model for Illustrious, as most Illustrious checkpoints tend to do a good job with male characters of different body types already. However, I found some models (especially the realism models) still seem to be mostly female biased, and generate female or futa type characters instead of males when prompting for more feminine or androgynous males.

So here we are again.

Before we begin, here are some quick notes.

This model, though it does a decent job expanding on the male body types, is still not perfect. Especially with femboys (mostly using realistic checkpoints) you may still get small breasts, this is usually best dealt with by adding 'flat chest' to the prompt. I've tried adding pectorals, and many 'male' specific tags to the prompt, as well as adding female, futa and breast specific tags to the negatives, but 'flat chest' in the positive prompt seems to be the best and easiest fix.

While the training data does include photos and realistic images, this model still works mostly with 2D styles. That said, it could potentially help boost male body types with other realism based loras, but alone you won't get much. I'll add some sample photorealistic images to a post to show the base quality, but I won't dive into it in this description.

V1 is good, and can handle most of body types fairly well, while V2 does a slightly better job at separating twink and toned, muscular and bara, as well as plump and chubby.

Now to the good stuff.

'Femboy', 'Androgynous' and 'Twink', oh my!

V1

V2

The core focus in all iterations of my Male Mix models was always to try and get a good feel for these three body types.

'Femboy' is the most feminine appearing male body type, usually with longer hair and feminine curves. Adding wide hips or curvy to a femboy prompt helps push this even further.

'Androgynous' males have feminine features (usually in the face), while still being a good mix of masculine and feminine overall.

And 'Twink' have a more youthful late-teens/early 20s look, that can go either way as far as masculinity and femininity are concerned.

(Yes, these probably aren't 100% accurate to the actual body types, but it's just a simplified anchor point using tags that already exist within the common captioning systems, and you are free to mix-and-match them to get whatever you want.)

Next we have the more general body shapes, 'Skinny', 'Toned', 'Muscular', 'Bara', 'Plump', and 'Chubby'.

V1

V2

These are common body types that already exist within most, if not all, current checkpoints to some degree.

'Skinny' is self explanatory really, and, while using words like thin and slim can sometimes get you a similar look, but from my testing skinny seems to work best for very thin characters.

'Toned', 'Muscular', and 'Bara' have a similar effect, giving more muscle definition to the character.

'Toned' usually gives a slimmer but still somewhat muscular physique or can be used to add muscle definition to characters, words like lithe have also been used in some models for a similar purpose.

'Muscular' and 'Bara' are often used together, or interchangeably, so they tend to have a lot of crossover, even with this model. But generally muscular keeps more of that "triangular" torso shape, while bara gives the more "square" or barrel-chested shape, mostly seen around the waistline.

Finally, 'Plump', and 'Chubby' also have a very similar output, with plump usually having a little more muscle definition, and chubby having more round features.

Again, these are commonly tagged (auto-tagging/booru tags) body types that most models have built in already, so this lora isn't necessary to capture them. My goal is just to try and push them a bit further, and provide better distinction between them.

Though, as you can see in the example images, even this model still needs some work. So I have plans for at least a v3 in the future with an updated, and higher quality dataset.

Until then, I hope this model provides some help and unique possibilities to anyone who uses it.

This model was trained using the Illustrious-XL v1.0 checkpoint, but all generations are made using WAI-NSFW-illustrious-SDXL v12 and Fooocus.

All of my models may be used for your own training and/or merges.