Note: The sample images are not representative of the specific style, as these models should adapt to various styles based on the checkpoint and other Loras used. The goal of these models is to capture the various male body types.
Male Mix XL
This is my attempt at making a more versatile model for producing images of men with various body types. It's not perfect, but it works.
The dataset contained art pieces of different styles and levels of detail, from anime to detailed paintings.
The main tag word used in training was simply "male", but other tags include femboy, androgynous, twink, toned and muscular. While it wasn't trained on words like thin, skinny, stocky, chubby, etc. it can still produce these body types.
For v3, I used a new and larger dataset, this allows the model to more accurately capture the various body types.
v4 Uses a completely new data set, which was based on my Illustrious checkpoint merge, Zabadu's Mix. This data set was smaller, but included images for more body types, including "Bara", "Plump" and "Chubby". For this version I also did my best to expand on some basic skin tones in the captions, so you may have to specifically prompt "light skin" for the basic lighter skin tone.
v1 and v2
v3
v4
+ 'Bara', 'Plump', and 'Chubby' body types
'Plump' and 'Chubby' don't have as much separation in SDXL it seems, so both tags give roughly the same result.
The various body type changes are more drastic depending on how you make the prompt, these examples are just using simple prompting.
With some checkpoints and seeds Femboy, Adrogynous and Twink tags may lose some of the body shape, and look more like a regular male body with only the faces looking more feminine or younger. I tried to fix this, but it probably comes down to just better prompting.
Though the v1-2 dataset didn't contain any photos, or hyper realistic images, the loras can create these to some extent. But they will most likely come out looking plasticy, or super smooth. I haven't fully tested this with photo specific checkpoints, or other photo loras, so I cannot say if those would improve the results.
For v3-4, I did add photo/realistic images to the dataset, so it should be able to create better images in this regard.
Using the SDXL refiner model may be necessary if you want photorealism, but can also cause more photorealistic results with other art styles.
Here's an example just using the base SDXL model and no other loras,
v1 and v2
v3
v4
+ 'Bara', 'Plump', and 'Chubby'
Again, for 'Plump' and 'Chubby' there is just too much crossover in the base model to see much difference here.
All models were trained using the SDXL v1.0 VAE Fix checkpoint, and all art style generations used the same checkpoint without the refiner, all photorealistic images used the SDXL refiner.
The v1-3 images were made in Automatic1111, while the v4 images were made in Fooocus.
All of my models may be used for other training and merges.