A working SDXL anime + ControlNet OpenPose setup, including the two model files that most people get stuck on because they are not reachable from the usual download page.
NoobAI-XL as the checkpoint, xinsir's OpenPose SDXL controlnet, and DWPose for skeleton extraction. 832×1248, 28 steps, strength 0.9, about 30 seconds per image on an 8 GB laptop card.
WHAT ACTUALLY BREAKS PEOPLE
· DWPreprocessor downloads yolox_l.onnx (217 MB) and dw-ll_ucoco_384.onnx (134 MB) from HuggingFace on first run. Without access to it the node dies with LocalEntryNotFoundError after a long wait. The README lists exactly where to put both files.
· Installing comfyui_controlnet_aux's requirements while ComfyUI is running fails with WinError 5 on cv2.pyd. Stop ComfyUI first.
· The pose source has to be a real human photo. An articulated drawing mannequin produces a garbled skeleton and the result only matches halfway — photoreal images generated by another model work fine.
The workflow also saves the extracted skeleton next to the render, so you can see whether the detector understood your reference before blaming the sampler.