Sorry the sweep is exactly 0 - 22.05 kHz so full bandwidth of the sampling rate.
There are always images at multiples of the sampling rate. Higher the rate, higher those images go in frequency and further apart from each other. Sampling rate is like a mirror, there are positive frequencies above it, and negative frequencies (inverse spectrum) below it.
So in the first picture there are images around every multiple of 44.1 kHz, and since content has 0 - 22.05 kHz bandwidth the images are completely adjacent due to no oversampling. To make the spectrum clean, analog filter would need to be able to cut out everything precisely above 22.05 kHz but nothing below. This is practically impossible task for analog filter, but not for digital filter. So the first images are really strong.
In the second picture you can see images around every multiple of 384 kHz, and since the content still has 0 - 22.05 kHz bandwidth, there’s plenty of empty space between the content and first image and the first image is already -50 dB down and much higher up in frequency.
From both pictures, we can see that with this given analog filter the images would be practically removed by having upsampling to about 3.2 MHz sampling rate. But now with the sampling rate limited to 384 kHz there are still some left, albeit total power of the images being much much smaller than without upsampling.
Now it would be already feasible to design a higher order analog filter that would already significantly lower the remaining images around 384 kHz multiples.
You can also see that upsampling significantly reduces the HF roll-off in the 20 kHz band.
Here’s the NOS mode output in narrow band:
And here’s the same with upsampling to 384k:
From the upsampled plot you can also see how the DAC’s THD increases as function of frequency (the increasing slope above 22.05 kHz).