Introduction to Understanding Perceptions of Real vs. AI Images
The realism of images produced by generative AI platforms is incredible. In fact, some images can’t be differentiated from those that are real.
In this blog post I’ve created a series of images, each of which contains a real image and an AI-generated image produced in Midjourney. The left image is titled A, while the right image is titled B. Also, a description of each pair is noted below the two images so viewers can confirm which one is real vs. fake according to the key at the bottom of this article.
Our effort should not be considered a highly-detailed analysis of how real AI-generated art appears. This is an extremely simple introduction to the subject matter, nothing more.
Real vs. AI Image Comparison Pair 1: Lighter Skin Tones
One of the images is, of course, AI-generated. This skin tone appeared to be the most common for a phrase like “photorealistic closeup of female face.”
Real vs. AI Image Comparison Pair 2: Darker Skin Tones
Next I wanted to try a different skin tone to see if that made any difference on realism. Given the enormous variety of people in the world, this could be the subject of an extensive analysis all its own.
Real vs. AI Image Comparison Pair 3: Landscape
Human faces are difficult to fake, but landscapes appear closer to reality in many cases.
Real vs. AI Image Comparison Pair 4: Seascape
This may be a tough one for a lot of viewers. Each of these seascape images looks fantastical, which is exactly how sunsets over the ocean can appear.
Real vs. AI Image Comparison Pair 5: Pizza
Wow. I think a lot of viewers will get this one wrong.
Real vs. AI Image Comparison Pair 6: Sunflowers
Which flower is real? Which flower is AI-generated? It’s extremely tough to tell the difference.
Real vs. AI Image Comparison Pair 7: Parrots
This one may have an indicator or two which imply the fake, but it might be tough to discern.
Results
- Lighter Skin Tones: A = Real | B = AI
- Darker Skin Tones: A = AI | B = Real
- Landscape: A = Real | B = AI
- Seascape: A = AI | B = Real
- Pizza: A = Real | B = AI
- Sunflower: A = Real | B = AI
- Parrot: A = AI | B = Real
Conclusions
All of the non-face objects were more difficult to differentiate than the faces. Still, AI faces are quite close to those of real people.
Perhaps we pay particular attention to the tiny differences and structural elements in faces more than other visual objects. If so, it could account for why faces are more difficult to mimic with generative AI than other types of images.
Special Thanks
I appreciate the work of the photographers of the images I used as “real” options, all of which were found on unsplash.com.
- Image 1 = https://unsplash.com/photos/T5AAr_PfdRA
- Image 2 = https://unsplash.com/photos/2OhHpRu2Fhg
- Image 3 = https://unsplash.com/photos/0TMJFpjlEFg
- Image 4 = https://unsplash.com/photos/jXn1Ua7OD38
- Image 5 = https://unsplash.com/photos/a66sGfOnnqQ
- Image 6 = https://unsplash.com/photos/3YIuhS0cmGs
- Image 7 = https://unsplash.com/photos/az2lhaX0s7g
Contact the Author
Please feel free to contact me. If you have a useful question or comment, I’ll add it here with my thoughts as appropriate, though I will never post your name or email address.
Bill Adams
[email protected]
If you would like to speak to someone at Streamline about a project, please contact us any time.
