The original startup simulator used cropped images of real people, and I could see in the git history where people had asked to have their photos removed, so for the remake I decided to use AI for the headshots. Each was generated with the grok-imagine-image-quality model. The prompt for each person had sixteen different attributes (gender, age, heritage, hair color, face shape, glasses, etc.), and no two images have the same values. ArcFace is then used to ensure each face has a minimum difference compared to all other faces. Anything too similar was kicked out and a new profile replaced it. The result was ~500 headshots each for the 2015 and 2026 eras. The gender and race of each image are tracked so that when one is selected, an appropriate name can be drawn for it. The list of last names was the largest table in the original generator. It seemed overkill for what it was doing, so for the remake I discarded all the names below a certain probability.
Given what we can do today with LLMs it would be even cooler if it thought of some actually good business ideas and created the website with mockups of them.
What did you use to generate human faces? As far as I remember, the earliest and most famous face generator ("This Person Does Not Exist") used GAN.
Anyway, my "Luciana Roberts" [0] has an artifact around her hair very similar to what older GANs often produced.
It's not a complaint, just a curious observation.
[0] https://restarted.io/?z=448760197027
The original startup simulator used cropped images of real people, and I could see in the git history where people had asked to have their photos removed, so for the remake I decided to use AI for the headshots. Each was generated with the grok-imagine-image-quality model. The prompt for each person had sixteen different attributes (gender, age, heritage, hair color, face shape, glasses, etc.), and no two images have the same values. ArcFace is then used to ensure each face has a minimum difference compared to all other faces. Anything too similar was kicked out and a new profile replaced it. The result was ~500 headshots each for the 2015 and 2026 eras. The gender and race of each image are tracked so that when one is selected, an appropriate name can be drawn for it. The list of last names was the largest table in the original generator. It seemed overkill for what it was doing, so for the remake I discarded all the names below a certain probability.
Cool, thanks for the detailed response!
> grok-imagine-image-quality
I find it curious that a 2026 model still occasionally makes same type of mistakes that 2018 StyleGAN made.
Given what we can do today with LLMs it would be even cooler if it thought of some actually good business ideas and created the website with mockups of them.
I legit thought that this was what was being sold here
This solution is less LLM and more parrots $h1tt1ng on newspaper. Keep clicking and see what else you find.
I got Shopify: https://restarted.io/?z=3119517161
wow, watch out, someone might steal that idea right out from under you
I'm sure it's one of those Thomas Edison/Joseph Swan situations. Keep clicking and see what else you find.