Inconsistencies caused by split-patch inference #55
Description
The current repository code performs patch-based inference on resized images during processing, then fuses them into a single large image using wavelet fusion.
However, this approach still results in a significant proportion of inconsistencies. For instance, when processing walls, some patches tend to smooth out wall textures during inference, while others generate more detail. This leads to the final image appearing as if multiple patches were stitched together. An example of such inconsistency is shown below:
This issue becomes particularly unacceptable when handling large-scale portraits, such as ID photos. I'm wondering if there are any effective solutions?