Pose follows only 'Twins' – is SDXL playing favorites on MPS?
Hi all, I’ve been testing various SDXL OpenPose ControlNet models on a Mac Studio (M2 Ultra) using ComfyUI with MPS enabled. Interestingly, I’ve found that the “twins” variant from xinsir/controlnet-openpose-sdxl-1.0 produces usable results, showing clear influence from pose conditioning. However, the standard version (diffusion_pytorch_model.safetensors) doesn’t seem to affect generation at all, even under identical conditions (pose input, prompt, resolution, etc.).
Does anyone have insight into why the twins model might work where the standard version fails? Are they trained differently for SDXL’s twin encoder architecture?
Also, if anyone has suggestions for other SDXL-compatible OpenPose models (or preprocessor tips) that have shown strong conditioning performance, I’d love to hear your recommendations. Also using A1111. Thanks!