This is not a Lora!

Well, ok... It is a lora, but I trained it on UltronMerge - a very "narrow" merge so you likely won't find it useful.

What is this then???

This a ready to train dataset! Yes, you heard it right. An actual factual 200 image strong dataset of husbando hunks!

A little bit of history...

I was training up a character lora for myself dearest and accidentally cooked up a very neat art style of male characters.

It's very likely that such style is promptable in some way, but I don't know the tags. So, I made a data set from that character lora I mentioned.

Since that data set is ENTIRELY synthetic (there isn't even an artist name in the prompt) I reckon it's reasonably safe to share it.

Is it good???

You be the judge.

I didn't go too wild with content - it's strictly ???? men with various zoom, rotation, hairstyles and colors.

As a freebie, I'm also supplying you with a matching set of tag files! They might be janky tags, but I wrote them myself! Not something you want to hear ever, I feel. But it's the thought that matters.

But seriously, I tried to keep it minimal and not come up with too many of my own tags, but you can always just remake them with AI tagger... if you no like :(

Am I allowed to use that dataset in California???

No idea. But I'm not stopping you. All images were created on my PC and I'm allowing you to do whatever you want with them.

Protweaks

There's obviously some things that can be changed to have a positive impact on dataset. If something comes to mind - I'll note it here.

- Tag eye color. You can tag eye color to have more control, otherwise it's mostly blue.

- Remove cropped images. Some images have top of the head cropped oh so slightly. You can filter them out to prevent AI learning to generate that.

- Mix in poses, clothes, backgrounds. The original lora I was making the dataset from was very limiting in terms of output, as it was a consistent character with one set of clothes. But using BaraGen lora and Ultronmerge you can diversify the dataset with clothes, scenes and poses if you'd like. Doesn't need to be a lot of images, as long as they're the same style - ai will easily learn that characters can be more than just nude.

- You can grab BaraGen lora and look up metadata for training settings. But take it with a grain of salt - the training data was made with the same model I later trained BaraGen on. It might require a bit more training to fully transfer the style to a different checkpoint. But judging by Kohya_ss previews, style trains within single epoch of 1 repeat (Kohya_ss previews are made with SD1.5 Base as far as I'm aware). If not, consider increasing 'Network Dimension' to 16 or to 32.

(BaraGen was trained on 20 epochs, just because)

BYE!

Good luck and have fun.