Unable to reproduce the same quality as the super resolution on https://satlas.allen.ai

Hello,
Thank you very much for making this project available, and kudos on the impressive quality of the super-resolution results displayed on the website.

I’m currently trying to evaluate your model to see if I can achieve results comparable in quality to the super-resolution on https://satlas.allen.ai.
If successful, the goal would be to fine-tune or retrain the model on a large corpus of high-resolution satellite imagery of Europe that we own.

At the moment, I’m unable to reproduce results similar to those on the site. My outputs have poor colorimetry, buildings/roads/parking lots lack coherence and are excessively subdivided into small elements, and trees appear scattered everywhere.

Here is my test workflow, do you see anything wrong with it?

1 - download multiple version of a sentinel2 tile:
```
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2024/02/17/S2B_MSIL2A_20240217T110019_N0510_R094_T31UDP_20240217T140755.SAFE/GRANULE/L2A_T31UDP_A036299_20240217T110019/IMG_DATA/R10m/T31UDP_20240217T110019_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/12/16/S2B_MSIL2A_20231216T105349_N0510_R051_T31UDP_20231220T034255.SAFE/GRANULE/L2A_T31UDP_A035398_20231216T105350/IMG_DATA/R10m/T31UDP_20231216T105349_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/10/10/S2B_MSIL2A_20231010T105859_N0509_R094_T31UDP_20231010T191518.SAFE/GRANULE/L2A_T31UDP_A034440_20231010T110442/IMG_DATA/R10m/T31UDP_20231010T105859_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/10/07/S2B_MSIL2A_20231007T104829_N0510_R051_T31UDP_20241101T152557.SAFE/GRANULE/L2A_T31UDP_A034397_20231007T105344/IMG_DATA/R10m/T31UDP_20231007T104829_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/10/07/S2B_MSIL2A_20231007T104829_N0509_R051_T31UDP_20231007T161323.SAFE/GRANULE/L2A_T31UDP_A034397_20231007T105344/IMG_DATA/R10m/T31UDP_20231007T104829_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/09/15/S2A_MSIL2A_20230915T105701_N0510_R094_T31UDP_20241105T020528.SAFE/GRANULE/L2A_T31UDP_A042991_20230915T105833/IMG_DATA/R10m/T31UDP_20230915T105701_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/09/15/S2A_MSIL2A_20230915T105701_N0509_R094_T31UDP_20230915T174938.SAFE/GRANULE/L2A_T31UDP_A042991_20230915T105833/IMG_DATA/R10m/T31UDP_20230915T105701_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/09/10/S2B_MSIL2A_20230910T105629_N0510_R094_T31UDP_20241104T131834.SAFE/GRANULE/L2A_T31UDP_A034011_20230910T110647/IMG_DATA/R10m/T31UDP_20230910T105629_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/09/10/S2B_MSIL2A_20230910T105629_N0509_R094_T31UDP_20230910T182201.SAFE/GRANULE/L2A_T31UDP_A034011_20230910T110647/IMG_DATA/R10m/T31UDP_20230910T105629_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/09/07/S2B_MSIL2A_20230907T104629_N0510_R051_T31UDP_20241026T083843.SAFE/GRANULE/L2A_T31UDP_A033968_20230907T105113/IMG_DATA/R10m/T31UDP_20230907T104629_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/09/07/S2B_MSIL2A_20230907T104629_N0509_R051_T31UDP_20230907T205611.SAFE/GRANULE/L2A_T31UDP_A033968_20230907T105113/IMG_DATA/R10m/T31UDP_20230907T104629_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/08/21/S2B_MSIL2A_20230821T105629_N0510_R094_T31UDP_20241024T134553.SAFE/GRANULE/L2A_T31UDP_A033725_20230821T110322/IMG_DATA/R10m/T31UDP_20230821T105629_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/08/21/S2B_MSIL2A_20230821T105629_N0509_R094_T31UDP_20230821T182038.SAFE/GRANULE/L2A_T31UDP_A033725_20230821T110322/IMG_DATA/R10m/T31UDP_20230821T105629_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/07/07/S2A_MSIL2A_20230707T105621_N0509_R094_T31UDP_20230707T194313.SAFE/GRANULE/L2A_T31UDP_A041990_20230707T105914/IMG_DATA/R10m/T31UDP_20230707T105621_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/24/S2A_MSIL2A_20230624T104621_N0510_R051_T31UDP_20240912T071700.SAFE/GRANULE/L2A_T31UDP_A041804_20230624T104623/IMG_DATA/R10m/T31UDP_20230624T104621_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/24/S2A_MSIL2A_20230624T104621_N0509_R051_T31UDP_20230624T183709.SAFE/GRANULE/L2A_T31UDP_A041804_20230624T104623/IMG_DATA/R10m/T31UDP_20230624T104621_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/14/S2A_MSIL2A_20230614T105031_N0510_R051_T31UDP_20240910T000108.SAFE/GRANULE/L2A_T31UDP_A041661_20230614T105029/IMG_DATA/R10m/T31UDP_20230614T105031_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/14/S2A_MSIL2A_20230614T105031_N0509_R051_T31UDP_20230614T182836.SAFE/GRANULE/L2A_T31UDP_A041661_20230614T105029/IMG_DATA/R10m/T31UDP_20230614T105031_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/07/S2A_MSIL2A_20230607T105621_N0510_R094_T31UDP_20241009T210215.SAFE/GRANULE/L2A_T31UDP_A041561_20230607T105905/IMG_DATA/R10m/T31UDP_20230607T105621_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/07/S2A_MSIL2A_20230607T105621_N0509_R094_T31UDP_20230607T183221.SAFE/GRANULE/L2A_T31UDP_A041561_20230607T105905/IMG_DATA/R10m/T31UDP_20230607T105621_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/04/S2A_MSIL2A_20230604T104621_N0510_R051_T31UDP_20240910T071227.SAFE/GRANULE/L2A_T31UDP_A041518_20230604T104623/IMG_DATA/R10m/T31UDP_20230604T104621_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/04/S2A_MSIL2A_20230604T104621_N0509_R051_T31UDP_20230604T184514.SAFE/GRANULE/L2A_T31UDP_A041518_20230604T104623/IMG_DATA/R10m/T31UDP_20230604T104621_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/02/S2B_MSIL2A_20230602T105629_N0510_R094_T31UDP_20240911T010411.SAFE/GRANULE/L2A_T31UDP_A032581_20230602T105829/IMG_DATA/R10m/T31UDP_20230602T105629_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/06/02/S2B_MSIL2A_20230602T105629_N0509_R094_T31UDP_20230602T153338.SAFE/GRANULE/L2A_T31UDP_A032581_20230602T105829/IMG_DATA/R10m/T31UDP_20230602T105629_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/05/28/S2A_MSIL2A_20230528T105621_N0510_R094_T31UDP_20240908T090134.SAFE/GRANULE/L2A_T31UDP_A041418_20230528T105909/IMG_DATA/R10m/T31UDP_20230528T105621_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/05/28/S2A_MSIL2A_20230528T105621_N0509_R094_T31UDP_20230528T185324.SAFE/GRANULE/L2A_T31UDP_A041418_20230528T105909/IMG_DATA/R10m/T31UDP_20230528T105621_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/05/25/S2A_MSIL2A_20230525T105031_N0510_R051_T31UDP_20240907T100603.SAFE/GRANULE/L2A_T31UDP_A041375_20230525T105028/IMG_DATA/R10m/T31UDP_20230525T105031_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/05/25/S2A_MSIL2A_20230525T105031_N0509_R051_T31UDP_20230525T173333.SAFE/GRANULE/L2A_T31UDP_A041375_20230525T105028/IMG_DATA/R10m/T31UDP_20230525T105031_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/04/20/S2B_MSIL2A_20230420T104619_N0510_R051_T31UDP_20240901T053820.SAFE/GRANULE/L2A_T31UDP_A031966_20230420T104622/IMG_DATA/R10m/T31UDP_20230420T104619_TCI_10m.tif
/vsicurl/https://sentinel2l2a01.blob.core.windows.net/sentinel2-l2/31/U/DP/2023/04/20/S2B_MSIL2A_20230420T104619_N0509_R051_T31UDP_20230420T160400.SAFE/GRANULE/L2A_T31UDP_A031966_20230420T104622/IMG_DATA/R10m/T31UDP_20230420T104619_TCI_10m.tif
```

2 - select manually the 8 tiles that have the fewest black pixels:
```
T31UDP_20231216T105349_TCI_10m.tif
T31UDP_20230420T104619_TCI_10m.tif
T31UDP_20230525T105031_TCI_10m.tif
T31UDP_20230604T104621_TCI_10m.tif
T31UDP_20230614T105031_TCI_10m.tif
T31UDP_20230624T104621_TCI_10m.tif
T31UDP_20230907T104629_TCI_10m.tif
T31UDP_20231007T104829_TCI_10m.tif
```

3 - reproject to EPSG:3857 (9.555mx9.555m) using GDAL:
```
gdalwarp -t_srs EPSG:3857 -tr 9.555 9.555 -r cubic -multi -wo NUM_THREADS=ALL_CPUS -co TILED=YES input.tif output.tif
```


4- inference
Use a slightly modified `ssr.infer` with `esrgan_8S2.pth` to ingest `(8 × 32, 32, 3)` patch images built by tiling the reprojected data, then stitch them.

Here is example of the image patches fed to `ssr.infer`:

Here is the result and what you have on satlas website:


I also tested with `esrgan_4S2.pth`, but not with higher `n_lr_images`, since the supplementary paper indicates that this only brings small gains.

As you can see in the screenshots, colorimetry is completely wrong and appears yellowish (although in my tests, colorimetry seems better with `esrgan_4S2.pth`).
Another major issue is structural: most buildings turn into vegetation, and everything looks “blobby.”

Questions:
- Do you see any problem in my workflow?
- Any idea how I could troubleshoot this?

Thanks a lot in advance,
Best regards,
Joachim