inference over sea water

Hi,
Thank you for your great job and making all the satlas project open source.
We are not able to produce super resolution images with the same quality as you display in https://satlas.allen.ai/map and would be happy to have your support.

We use 8 sentinel2 L1C images, wrapped to 9.555m resolution + bilinear. We use the "esrgan_20m.pth" pretrained weights, 9 bands ["TCI", "B08", "B05", "B06", "B07", "B11", "B12" ]. For the non TCI bands, colors are converted from UInt16 to [0-255] range deviding by 8160 and clipping to [0 - 255].
We generate TIF tiles 32x256 pixels / 9 bands to provide to "infer.py" or "infer_grid.py" (modified to open tif instead of PNG).
The generated super resolution is good on land. We have noted the generated roof tops are often more red than in your web map. But on sea, the generated images have almost no sea colors transparency, sea bottom is generally not visible, at the opposite on you web map. We have tried to replace TCI bands by B04 B03 B02 without success.

Can you describe in details how you generate the png tiles from the JP2 sentinel2 sources.
A kind of unit test example would be helpful : you provide the 8 sentinel2 source images, the script to reproject (it is already on the github) and the reprojected images, the script to generate some PNG tiles and the reprojected images, then the generated super resolution images. Like that we would be able to understand if we are able to produce exactly same results and find any error in our script. The example must cover sea water, with water transparency.

Below image from satlas.allen.ai/map

Below our generated image. Dredged channel is not visible.