Enthusiast Weekly Update #7 (Pixiv Fanbox)

Something a little bit different for you today!

The past week or so has been pretty exciting in the Stable Diffusion community. ControlNet has been making waves, and is looking to bring yet another revolution in AI image generation. But what is ControlNet exactly? I'll try to give a brief explanation.

Say you have a pose that you really like. For this example, here's the source image I'm using:

Not very inspiring, but ControlNet can take this image and extract the pose from it using OpenPose. There are a few other ways to extract metadata from the image, but so far OpenPose has been the most useful to me, as it only picks out the character and ignores any background details:

Then, on the basis of this pose data, a good starting prompt, and a little bit of experimentation you can generate essentially endless variations on the same pose. All of these are straight from txt2img, without inpainting modification:

Loli, sitting on the floor, naked

Loli, lying in bed, underwear

Loli, dark skin, evil grin, pirate, sitting by the ocean, bikini

Milf, petite, qipao, sitting on the sidewalk, flashing panties

Girl, lying in grass, naked, small breasts

Pretty wild, huh?

Alright, I'll add one more image. I've been practicing on doing some established characters, because I would like to get into commissioned character art a bit more. I made this for a Pixiv follower, the asymmetrical details were quite a nice challenge:

See you next week!