I tried it out with a picture of Brad Pitt. The results are NOT close, unless you think "Yes that's a white male" is close. The hair color was wrong, but that could have been due to using a picture of Pitt where his hair is dyed blond but his black roots or showing, or just due to shadow.
I guess they have many hundreds or thousands of images pre-generated, and when you upload a photo, they find the closest ones.
It definitely has some fitting issues; Brad Pitt is a good example as he dyes his hair. I've googled his hair situation so you don't have to. He most probably has gray hair, dyed dark, then highlighted.
That's a challenge huh! As a result, you can feel some of these guys manipulated hair color: https://i.imgur.com/bC6AxXO.jpg
Anyway, I had better luck with Angelina Jolie. Hairwise, her hair looks also toned, but our dataset has way more colored women:
https://i.imgur.com/WmTzRAZ.jpg
Ya the bottom left is pretty good! I've actually bought from the site before so please don't think I was knocking it too hard. It's very useful when you need a unique image of a face.
Thank you for buying and thank you for your encouragement!
My personal fault was positioning. We can’t make everyone looking similar. As you said, we have a bunch of files on a disk, heavy indexed ones. We spend a year working on diversity, including running as photo studio, but it has limits yet.
So, we needed to invent a use for it. Anonymizing seemed like the opposite of resemblance, yet people wait for the latter. More resemblance (and prettier the face) more they like it.
I wonder if the current tech is marketable without false expectations.
I like the idea! That would be a cool next step after this mvp!
We could either pre-generate (see the first comment - it’s an accurate guess), or generate just in time.
The first is way easier - same architecture, more storage (we are moving from instance’s storage to S3 so it will be an easy step).
On-demand generation would be much more expensive as you have to have a GPU server ready, and queue requests. We have it with https://icons8.com/upscaler ; it’s a bit of work and expense.
Additionally, you have to make it a paid service. Any ideas who could pay for that? Our usual customers are https://generated.photos/use-cases
On the boring side, it’s probably not worth it. Generated photography is detectable. A trained eye could spot it; big social media have models trained to a lot it.
Scraping a real life dataset makes more sense to me. Moreover, there are tons of ready ones.
I guess they have many hundreds or thousands of images pre-generated, and when you upload a photo, they find the closest ones.