As title; currently our approach to tiling is to circularize every 2D convolution in the model regardless of size or purpose, on the assumption that every 2D structure represents image data in a spatial layout.
This does generally seem to work fine, but it is possible that some distortion is being caused by circularizing convolutions that are operating on other kinds of data. Resolving this issue would require going through the SD model, understanding the purpose of each use of Conv2D, then modifying this script so that it only circularizes convolutions that genuinely represent image-space data.
As title; currently our approach to tiling is to circularize every 2D convolution in the model regardless of size or purpose, on the assumption that every 2D structure represents image data in a spatial layout.
This does generally seem to work fine, but it is possible that some distortion is being caused by circularizing convolutions that are operating on other kinds of data. Resolving this issue would require going through the SD model, understanding the purpose of each use of Conv2D, then modifying this script so that it only circularizes convolutions that genuinely represent image-space data.