Openpose sd v1.5 weights different from original sd v1.5 weights

[wenquanlu]

Hi, I downloaded the controlnet weights control_sd15_openpose.pth for openpose sd v1.5 from this site https://huggingface.co/lllyasviel/ControlNet

However, after I extracted the stable diffusion weights from this file, I compared it with the original sd v1.5 weight released in https://huggingface.co/runwayml/stable-diffusion-v1-5

I found the weights for encoding and middle block are the same, but the decoding blocks are different. So I guess it was trained with sd_locked = False? Can you provide the openpose weights in which the original sd is completely locked by any chance?

Furthermore, seems like in HuggingFace diffuser ControlNet implementation, they forcebly merge the extracted pure controlnet weight and the original runwayml/stable-diffusion-v1-5 sd weight together, does this suggest we can also merge this without degrading the generation quality much?

Thanks!!!

[lllyasviel]

Hello cnet 1.0 has an sd_locked option, see also "sd_locked " in https://github.com/lllyasviel/ControlNet/blob/main/docs/train.md

but it turns out that such unlock does not have influence to the results, and the sd_locked option is removed in cnet 1.1

See also https://github.com/lllyasviel/ControlNet-v1-1-nightly and the models https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main

[wenquanlu]

OK, cool, thank you so much!