I wasn’t satisfied with the first result, so I kept trying. First I tried the SoftEdge ControlNet. That was a lot harder to work with and didn’t produce good results. The breakthrough came when I tried using Latent Couple instead of Regional Prompter. It seems better suited to precision work like this. Next I’m going to try using just the OpenPose ControlNet to keep her in the same pose so that the Latent Couple mask is always in the right place. That won’t require using GIMP to edit the Canny input.

Aside from the edited Canny input image, this just required one step, so I’ll share the generation info. It’s missing the Latent Couple config unfortunately.

a portrait of beautiful naked woman laying on a bed with her legs spread, beautiful face, long brown hair, white sheets, portrait, perfect lighting
AND (mouth vagina:1.4)
Negative prompt: painting, drawing, (vagina:1), pubic hair, nostrils, clitoris
Steps: 20, Sampler: Euler a, CFG scale: 9, Seed: 2820077928, Size: 512x768, Model hash: 9584e2c050, Model: cyberrealistic_v32, ENSD: 31337, ControlNet: "preprocessor: none, model: controlnetPreTrained_cannyV10 [e3fe7712], weight: 1, starting/ending: (0, 1), resize mode: Crop and Resize, pixel perfect: False, control mode: Balanced, preprocessor params: (512, 100, 200)", CFG Rescale phi: 0, Version: v1.5.1-1-g0c72f821