This model represents my first attempt at creating a LORA. Specifically, I trained it on my Illustrious images to see if I could get those qualities into Anima. In that goal, I mostly failed.

However, I still find what the LORA does do to be wildly appealing, so I'm posting it here anyways so I can tag it in pictures I make. So what does this version do?

• Changes Jawlines and proportions to give a more mature and (sometimes) a bit more androgynous look

• Changes the overall shape of "Muscular" to more bulk to futanari, women and even men

• Makes penises thicker without adding much length

• WILDLY improves crotch bulges across the board

• And various other small touches relating back to the source images

Issues:

• I used a lot of ControlNets in Illustrious and if a composition gets too close to one it will "snap" to that pose instead of what the checkpoint wanted

• Tends to make images more sepia toned

• Breaks literally every POV scene I've tried it on (despite POV being in the training data)

• Can give characters uneven skintones especially at higher weights

• Turns Fembois into full-throated men or futanari depending on how they're prompted

• Picked up some bad habits from Illustrious at higher weights, like hiding hands.

• Struggles with scenes of 4 or more subjects

I've done a lot of testing of this model in the interest of my own learning. If you want to use this LORA, here's some recommendations:

• Weights vary by Checkpoints:

  • For Mokubapoint v3 and other "shaded anime" checkpoints, best results are usually around .8 and .6

  • For Terra Rising and other "line art" checkpoints, the best results are around .6 to .4

  • For RealDream Turbo and other realistic checkpoints, best results are around .3 to .2

  • On any checkpoints, at low weights, it tends to make characters look more mature, muscular and hung (if they have something to hang)

• Positive prompting:

  • Muscular, thick penis, stretched pussy, visible penis base (for penetration) and punk hair are all key tags that will have large effects

  • I recommend for negative:

    • sepiatone filter, sepia, uneven skintone, long penis, darker penis,

I clearly need to pick and generate my dataset better to try to resolve the issues I've found and make something that better represents the style I was going for. That said, I'm also hoping to make a better version of this model that does what this does as well.