Upsampling - is it improving the sound?

You say you are using the same mix and switching the upsampling on and off with this singular mix, correct? You need to mix the upsampled sound, not mix the un-upsampled sound then switch it on. I would assume a better sound WOULD have brighter and clearer sound etc. but when mixed in properly this should result in a better overall sound. I think if you create a separate mix for each state of the upsampling, on and off, you will have a more accurate sense of what the upsampling does.

Just curious, how do you define "overall quality"? Why would high end content be immune to change with a quality change? I mean there should be differences, right?

Thank you guys, I think I need to enable oversampling on playback before starting mixing or begin recording everything in 96k :)

Usually oversampling do the better job than upsampling.
Oversampling in rational integer multiples like 2x, 4x, 8x or even 128x with pure sine algorithm like delta-sigma in ADC can help us to reduce aliasing distortions and in DAC can help us to reduce distortions from digital filters (or pull the distortions far away to above 100kHz) thus less sound details lost.
Upsampling could do the similar great job with really excellent pure/clean algorithms. 
 
Applying oversampling/upsampling in digital to digital stage like PCM sampling rate conversions in various DSP or softwares like SONAR or sound forge could help reducing similar distortions as well, especially very good in reducing aliasing, but again, a great algorithm is the most important. And it's not always necessary, the most necessary conditions would be some physical-modeling synth like what Cakewalk told us, or high-quality Reverb effects like 2CAudio Aether.
 
Resampling/oversampling jobs may also get worse than original audio, especially result in more and more aliasing (maybe that's the "more digital" you mentioned). So to get the best resampling/oversampling resulting in as less aliasing as possible, some following parameters should be kept an eye on. Like the resampling in my Sound Forge by Izotope:

Steepness should be the steepest, unless if you wanna hear more and more aliasing noises, why you still have to keep any high freq. sounds which higher than the half of your new sampling rate?
Max. filter length should be the longest, to get more and more precise result.
And the most important: Alias suppression, should be the max., unless if you love those aliasing noises.....
 
And finally, all things should be done in 64bit FP. That's why I always process my audio works in full 64bit rendering.

 
 
I'm not sure if the resampling quality in SONAR is as good as the Izotope in Sound Forge one. If yes, I think Cakewalk should better let us adjust the resampling parameters like above.
 
Remember oversampling/upsampling does not give you more extra sound quality or extra real sound details, it does not increase your sound quality, BUT avoids from losing any sound quality and sound details, which makes you lose as less as possible, which far less than not doing oversampling/upsampling. And again, all things should be done in full path of 64bit FP process.

Anderton

parco
I'm not sure if the resampling quality in SONAR is as good as the Izotope in Sound Forge one. If yes, I think Cakewalk should better let us adjust the resampling parameters like above.

 
SONAR's is among the best. Check out sample rate conversion tests for various DAWs and editors. You might be surprised...a lot of programs used to be pretty horrible but have improved. SONAR's has been excellent for quite some time.
 

Remember oversampling/upsampling does not give you more extra sound quality or extra real sound details, it does not increase your sound quality

 
Actually it does for processes that occur "in the box," like amp sims and virtual instruments, because the higher sample rate eliminates audible foldover distortion. Please listen to the video in post #7.
 

The Sonar X3 one you mean? well......... huh.......... still not bad....... although not as good as the Izotope 64bit SRC Steep No Alias one............ 

I actually submitted them the samples for X3, I think it's been the same since Sonar 6 but they didn't have a "recent" version on there and there's always the chance people will assume a different program version will have a different conversion algorithm. I figure X3 is still recent enough though, and last time it took them months to upload the results, they probably have better things to do.