If you like what I do please consider supporting me on Patreon and contributing your ideas to my future projects!
Poses to use in OpenPose ControlNet modeled from professional nude model photographs.
I tried to pick the best ones that gave the best results. If your model looks mildly contorted it might be because it wants to show the front side when the pose was taken from the backside. Typing "show ass" and "view from back" into the prompt will usually fix that.
If your model is clipping outside of your image it might be because you are using a different aspect ratio then what was used to make the pose. Try using a similar aspect ratio. 768 width x 512 height works well.
If the pose isn't coming out the way you want try lowering the Ending Control Step.
For #37 you can type "dildo, vaginal insertion" into the prompt.
For #63 you can type "licking ass, cunnilingus" otherwise the other girl might rest her head on the girls ass like a pillow.
For #80 you can type "lesbian, yuri, kissing".
For #88 you can type "1woman, 1man, blowjob, penis, fellatio, man and woman should both be in view". You can get more accurate results with regional prompting. This one isn't always the most accurate as the penis likes to float in the air sometimes.
I used Majicmix for testing.
Invite to my public discord: https://discord.gg/UFkPTfTDWs
Description
Added example images of every pose
FAQ
Comments (12)
The one single pose I can't find, or prompt, or create in img2img is The Pose, the actual best pose. The one where she is laying on her stomach, ass over shoulder, I don't care feet up or not. I can understand the difficulty from the front, but even off to the side (or isometric angle).
Laying on front is a pain in general, even though you should be able to prompt a standing back view controlnet pose into lying, somehow it seems to know and puts the figure upright. Lying poses just seem to be a problem in SD, and forget about sex positions, loras seem to be the only half solution to that.
I have only started with controlnet, and maybe I can combine with loras, or run dual conrtolnets for couples, but that all gets complex. For some things, I have tried running loras with img2img with limited success. If there is a better way for these things, let me know, thanks!
Use controlnet to generate a pose from an image like I did with these
@getphat In img2img setting, the face changes beyond recognition((
what needs to be done to only slightly change the style of the face and leave the characteristic features unchanged?
@proevgenii I use txt2img mostly, but you can try turning the denoise down
Hey send me a picture of exactly what you're talking about and I'll include it in my next upload
@getphat I'm just trying to generate some pictures of myself, so I'm using photo with my face in Img2Img, the quality and photorealism are cool, but it changes a person beyond recognition, but I would like the face to change its style, but retain its characteristic features / recognition
I tried to play with the parameters, but did not get the desired effect.
Is it even possible to make the model generate characters with my face without finetuning?
@proevgenii Try controlnet reference or roop
There aren't any JSON files in here...
Where Should i put the files?
inside ControlNet Model folder?
When you open control net, you put the image into that box. Not a folder, but in Stable itself.
Anyone get this working in Forge?
Controlnet Integrated is weird.
i wonder know too. It is hard to understand how them works. (openpose, aditional networks, inpainting and controlnet)