ONLY TESTED ON AOM2 HARD MODEL

Doggystyle LORA in the style of the danbooru artist 'tomubobu'

https://danbooru.donmai.us/posts?tags=tomu_%28tomubobu%29&z=2

Delivers more consistent doggystyle gens. Use at your own risk, requires some gacha.

I would recommend using '???', '??? from behind', 'standing ???', 'standing', 'doggystyle' separately rather than combining the tags as it leads to more limb gacha. I have been training this thing for 15 iterations now and I'm bored of it. It's not as good as it can be, but it's absurdly difficult to strike the balance between high enough epochs to replicate the positions without burning the LORA due to the artist's unique style.

Set to 0.4-1.0 strength (New version can work out better at higher strengths). Use tags 'jagged lines' and 'sketch' to replicate more of the artist's style.

Trained on:

3e-4 unet/lr

6e-5 text encoder

128 dim/alpha

Data set of 10 standing, 100 all fours images, set to 35/20 repeats respectively.