Evaluating how different keywords affect realism in SDXL

We tested the SDXL model with a base prompt and different keywords to see how they affect the realism of the generated images. We found that most keywords didn't significantly change the realism, but some did. Adding photographic equipment related keywords produced the most realistic images. Adding "photorealistic" and similar keywords did not improve the results.

In popular AI communities such as Civit AI or PromptHero, prompts for realistic photos often include keywords like "hyper-realistic," "photorealistic," and "depth of field." But do these terms really enhance realism?

To conduct the test we test 20 keywords, each time appending them to two base prompts: "A realistic photo of a man, upper body, color" and "A realistic photo of a woman, upper body, color". The base prompts alone produced realistic images, but they were obviously AI Generated, with the skin looking a bit too smooth and the eyes a bit too perfect.

Our goal was to find out which keywords produce images that make you question for a moment if they are AI generated or not. This criterion was the basis for our image ratings.

For each keyword we generated 8 photos, 4 for each base prompt, and rated them on a scale of 1 to 3, with 1 being obviously AI generated and 3 being indistinguishable from a real photo.

Each keyword belongs to one of these categories: cinematic effects (depth of field, shallow focus, etc.), name ( a woman named jennifer, a man named joey), detailed (hyper-realistic, photorealistic, etc) and photography equipment (nikon d4s, lomography xprochrome film, etc)