OpenPose and depth images for ControlNet.
Recommended prompts are the below.
2girls, from behind, from belowRecommended control weights are the below.
OpenPose:
0.8
I made these images by PoseMy.Art.
I strongly recommend to use with Regional Prompter.
Description
FAQ
Comments (4)
I do like these, but it's interesting, even with 'rear view' or 'back view', half the time models want to twist the figure around to show front anyway. I can see that the sides have different colors, I don't get why the model doesn't stick to it. Then at the higher controlnet strengths it messes with quality and even composition.
And don't even get me started on perspective, if you don't prompt 'above view' or 'low angle view' you have no chance regardless of the pose. A low angle pose without prompting will usually just make a front view figure with huge legs, a narrow upper body and tiny head.
My most consistent results have used a Lora with compatible pose and specific prompting together, but finding that magic combination is a process for sure.
Yeah, I agree it. I have seen twisting in some images using this OpenPose. Perhaps, I can reduce twisting by adjusting OpenPose, but it takes a long time...sorry. I also think that both prompt and OpenPose are important. It would be nice to change the control weight of ControlNet and change the "Control Mode" . Don't forget to use not only from above, from below but also close-up, ass focus, etcetc... .
I have a particular idea for a pose for 2girls but my loras keep getting mixed I'm new to this
Try using Latent mode. It would be nice to reduce the weight of Loras. Unfortunately, some Loras are very easily mixed up...