Rather than simply replacing the base unet of the sdxl model with refiner unet, why not directly make an extension that can schedule unet replacement according to the number of sampling steps? In this way, different unets can be combined at will, and even different versions of ckpt can complete the same picture together. To switch between different versions of ckpt, you only need to use the corresponding version of vae to re-encode and decode the image.
Rather than simply replacing the base unet of the sdxl model with refiner unet, why not directly make an extension that can schedule unet replacement according to the number of sampling steps? In this way, different unets can be combined at will, and even different versions of ckpt can complete the same picture together. To switch between different versions of ckpt, you only need to use the corresponding version of vae to re-encode and decode the image.