Trained on UltronForever. Compatibility with other checkpoints it not guaranteed.

Round 2!

New year is approaching and it's time for a second round of presents!

I really liked how styles lora turned out, so I trained some more in that direction.

Not much to add here. It the same but MORE.

Way MORE.

There're 7 styles this time. As such, it's exponentially harder to balance concepts. If you liked pins-v1, don't go windmill slaming it into recycle bin just yet, it could still be better in some ways than pins-v2.

Dataset was around 1k of randomest images, making output extremely varied between epochs. I'm uploading the last epoch, because it took in the most of training data "ideas", but it ended up a little high on contrast in exchange. If after making some previews, I'll feel an earlier epoch was unarguably better, I'll upload that as well.

Usage

1. Pick a style:

'photo art, man' - a bit of an afterthought, but I felt I could spice the lora up a little with some diverse AI generated hunk photos. You can do any amount of weight for this to diversify photo style of the main model.

'render art, man' - plastic 3d model looking style. Contained some reasonably famous image board art, so you might find it somewhat familiar.

'digital art, man' - general dramatic fantasy art. Digital part in this, was airbrushed look and 'too fine for human' details. Most cooked style, recommended Lora weight <0.75.

'illustration art, man' - very similar to digital art style in content, but instead details are more coarse and art medium texture is often visible.

'pencil art, man' - rough sketch/lineart style.

'niji art, man' - anime. This style is not native to UltronForever, that's why output quality varies greatly. Lora weight might need adjusting per prompt.

'concept art, man' - anime character reference/concept sheet style. Niche busy looking style popularized by generative AI. Similar to 'niji art', but still it's own thing.

('man' is still there, because it seems like, it was beneficial for the first lora, so I included it in second version as well)

2. Deepbooru tags. Even though, training data was tagged with Deepbooru tags, I feel they're optional. Lora works perfectly fine with any kind of prompts.

3. Concepts. Training data contained some repeating imagery that ended up trained as concepts. Here are some ideas to try:

bara, wings, fire, water, faceless, orc, fairy, elf, drow, orc, demon, devil, horns, colored (blue, grey, red, white, green, purple) skin , tattoo, elemental, cyborg, robot, glowing eyes

(place those near the start of the prompt and if it still doesn't appear, add extra weight to the tag; a word of warning - not every concept works equally well with every style, especially difficult ones like 'merman')

(honorary mentions - did not entirely get trained, but sometimes do work)

tiger, wolf, bear (anthropomorphics, very few images, will mostly rely on base model)

merman (fins and tail are all over the place, but can be inpainted reasonably well, and medium shot of a merman works quite often)

holding sword (it almost works, and even can appear good enough sometimes)

4. Latent upscale is highly recommended. Lora is slightly cooked, as such base output is often a bit fuzzy (shouldn't be news if you ever used UltronMerge).

5. Weight. I mostly use it at weight 1. But if some styles or even specific prompts appear cooked and oversaturated, try lowering lora weight to 0.75 or 0.5. How cooked imagery appears seems to vary greatly between prompts.

6. SFW. Dataset was still entirely without full frontal ??????. A few butts maybe.

Happy New Years!

Good luck and have fun!

How is everyone doing?

It's been a while... I haven't posted for a full day! Many of you must have been worried sick.

I'm glad to report, that I am well and good! Antis didn't get me just yet! I'm black belt in keyboard-fu after all... A little bit of 'gtfo', some 'no u', and they be droppin' like flies.

How is everyone's Christmas cheer?

I couldn't leave you all without a half-arsed self-made present. You know, nothing big, and um... A little flaccid, the usual.

What is it?

Do you guys, like, ever go to internet image boards? When you get there, you're always teased to death with amazing AI hunk images of unknown origin? No? Just me I guess...

But they teased me enough! Today is the day I say NO! And generate my own amazing hunks of unknown origin!

Welcome - Pins Lora!

Related to pinterest only in entirely coincidental name similarity (obviously), this Lora will provide you with not just one, but !4! unique art styles for generating handsome male imagery! Yes, you heard it right - 4 styles for the price of none! What a deal!

Woah!

As I trained it, I didn't intend it to be final, because I went for very coarse deepbooru tagging that often didn't even match the contents of the images. But it turned out quite good. We won't call it final for now, because there's place for improvement, but as is, it's already quite interesting.

Did only minimal testing as I generated previews. If any issues or usage suggestions arise, you will find them here, as usual.

How use?

- 4 main tags 'digital art, man', 'illustration art, man', 'pencil art, man' and 'niji art, man' (no idea why I felt the need to add 'man' there, and it might even work without it, but if it doesn't, use the full tag)

- Rest of tags are your typical deepbooru stuff like '1boy, male focus, bara' and so on and so forth. Use those at your own discretion.

- Lastly, add your prompt. It might work, it might not. You know how SD1.5 loras are.

- There were no ???? images in training data - shirtless or tastefully ???? with covered groin at most. So, it's only as pervy as the checkpoint generating images.

- Hands are weakish but work consistently in generic poses. Anything 'dynamic' needs to be fixed via outside tools.

- Styles are consistent as long as prompts are kept simple. Large amount of quality tags makes styles bleed amongst each other.

- Lora does work with any type of content (not just male characters) with highly varied degree of success (landscapes seem to work best, abstractions are also cool).

Merry Christmas!

Good luck and have fun!