inference speed

Hi, I tried your method in the huggingface demo and it works very well.
The current method when dealing with large resolutions will divide them into overlapping blocks of images of size 512*512 and then enter the model in turn. I would like to ask if you have made any improvements in this area?
Do you plan to open source new models and processes in the near future? Looking forward to it!