I see funny faces. Do you?
Using AI tools to visualise faces in everyday objects.
Have a brief look at the image below. What comes to your mind first?
A Ghost? A soldier with a cap? A French painter? I don’t know.
But I’m pretty sure you saw some face on that vase. There’s a name for this thing — Pareidolia! Simply put, it is the tendency of the human mind to perceive patterns, such as ‘faces’ on everyday objects or on cloud formations.
Since my mind has always been extra excited about anything that comes across it, I’ve been experiencing intense Pareidolia all my life. Eventually I started capturing them on digital photographs and kept sharing them on Instagram as @facetagger.
And ever since, I’ve been pondering the idea of visualising these faces that I see in these pictures. Creating doodles, digital paintings, photo collages and many more ideas went by.
Then it struck me! What if I use AI?
Well, I did exactly that. I asked the same question to AI — “What do you see?” And this was the result:
How cool is that!
As you can see, AI’s imagination was very vivid, and realistic, although I’m not sure whether it could accurately match the picture that popped up in my mind when I looked at the original.
I tried the same steps on a lot of other facetags as well :
I am so glad that the AI technology has been democratised so much so that even common folks like me can have fun with it with almost zero knowledge in coding. It was so much fun playing with these amazing AI tools available online.
Here’s how I created these images
At first, my idea was to run a machine learning model that could imagine faces on the fly. Interestingly there were a few methods I found online that could generate faces from sketches or annotations such as Pix2Pix. But as mentioned earlier, these were far beyond my abilities to understand them thoroughly and to run them successfully on a local machine.
Luckily I stumbled upon a product called Sketchface AI that looked promising. It was a product that was supposed to help artists create faces from line sketches.
Interestingly, it also had the ability to import real face images to edit them by drawing over it.
If you plan to try it, be warned that the results won’t be all that positive always. We humans have a very strong sense of face recognition, brought to that stage by millions of years of evolution. Even though it was very easy for us to perceive faces in those images, it wasn’t the same for machines.
This was probably because the initial conversion from image to sketch itself was not happening correctly.
But in some cases, I was able to get results that were good enough on the first try itself.
However, these generated images weren’t of high resolution, and they also lacked the realism that I was expecting.
That’s when I found an interesting photo restoration tool powered by GFP-GAN model as explained in great detail in Louis Bouchard’s blog. And the results were great.
Well, that’s for now.
If you too have intense pareidolia like me, check out more such stuff on my Instagram channel @facetagger
Endnotes:
I also experimented using Generative AI’s like ‘DALL·E’ and the results were mindblowing! Read more about it here.
A big thanks to Vineeth Nair for getting me get the SketfaceAI app; I owe you a big one.
