The open-weight model (shinuh/tf-restormer-baseline) produces 16kHz audio

[Sytronik]

I tested the official weight, but its output contains signal only under 8kHz as the following spectrogram.

• input
  • 
• output
  • 

But the official demo shows upsampling capability (to 48kHz sampling rate).
Is this just a bug? or the open-weight version is trained on 16kHz target?