I am a bit confused. You say you are a PhD student. At that level of education, it should be trivial for you to figure out a methodology to objectively judge photorealism of the images.
You simply gather large enough data set of synthetic computer generated images with varying levels of what an average person would considered “realism” (not obviously stylized), and a large enough data set of photographs of similar subjects/topics to the computer generated images data set.
Then you make a simple system which displays random images from both data sets, and for each image, asks people in a randomly selected test group whether they think given image is a photo or a computer generated image.
Once you have gathered data, you filter out the photographs, and then you sort the computer generated image results by the percentage of correct answers. If you sort it in an ascending order, then the computer generated images which were voted to be photographs the most (the incorrect answer) are the most realistic ones in the data set.
So the criteria is simple: How many percent of randomly selected people would confuse given computer generated image for a photo. The larger the percentage, the more photorealistic given image is.