The exam was to make a DL model that takes an image of 2 superimposed images as an input and gives the 2 separated images as an output and the grading depends on the metric which was the MSE loss (without pre-processing).

All I did was going through the blogs, github, anything I can get my hand on and use all models and use the model that yields the lowest MSE.

Sure some models were better than others but to a certain degree, they all worked. I mean it's not a classification problem so yeah, it should 'work'. But what I'm trying to say is they all worked at the end. I used models for image denoising, image segmentation, image resolution, etc and all of them worked.

Which made me wonder: for a problem such as that, what's the point of academics trying to improve models for tasks like that? Plus: why do they try so hard to compare each other when you're already at results good enough, you can't tell the difference with the naked eyes (a 0.0003 MSE vs 0.0004 MSE, for example)?

It was something I was just asking myself while doing the project (or searching for a code to copy/paste lol).

Btw I studied the theory and I get it, but it was just easier to do that - especially that the grade is stupidly inversely proportional to the MSE: the lower the MSE, the higher the grade.