Problems making sdxl inpainting models using modelmerge.

Whether you use the workflow in https://github.com/comfyanonymous/ComfyUI_examples/blob/master/model_merging/model_merging_inpaint.png. Or change to A+(B-C)*m: A=Official-inpainting; B=fine turning basemodel; C=Original base model.
The model that have been saved by CheckpointSave all have the same specs as the base model. (It does have the inpaint function). But not loaded with UnetLoader.
Whereas the inpaint model generated by auto1111webui has the same specs as the Official inpainting model and can be loaded with UnetLoader. However, using such generated inpainting model in comfyUI, the generated image is exactly the same as Official inpainting model.
So. Which inpainting model should I use in comfyUI?