Performance of SDXL-Refiner

[bee992]

Thanks to your great work! It looks so impressive.

Before you release the controlnet version, I had tried to keep the main content of image while upscaling. Because after the img2img, many details were always be changed. So I tried to replace the SDXL as SDXL-refiner.
Using images generated by SDXL as input, I had recevied a bad result of 20482048 image, while the reconstructed 10241024 image was good. I found that during the tile_decode of phase-2, the output patchs are all blur, what leads to this problem? Can you help me solve this problem, I think it will be fun if we can use SDXL-refiner to do real images upscale. Thank you.