Headshots are pretty vulnerable because no-one really cares if it's fake, as long as it actually looks like you. But there are a number of areas of photography that don't really have to worry about AI because the entire point is capturing the moment.
For example, wedding photography. A picture that looks like you and your husband dressed up doesn't cut it. It's the moment, capturing everyone together, the crowds, etc.
The best I can imagine is that you could have someone walking around with a basic phone (or CCTV) wide angle capturing everything in low res, and then an AI could take those images and create convincing looking upscaled versions of them with interesting perspectives and composition etc. I still think though that people will want the real deal.
There actually is a guy doing fake wedding photos for people who lost theirs (in fire, didn't have backups, etc) or didn't have the nice wedding they hoped for, and it's quite a successful business.
Another case, it's very popular in Asian cultures to go on a trip around Europe and get "wedding" photos in wedding dress in front of famous castles. There's no reason why these photos couldn't be generated with AI since the moment is fake already - and it's all about the photo. I worked for a company that was driving these people around, they spent max. 10 minutes at each location.
Sounds like the 90s and Glamour Shots all over again. The point of those photos was to look expensive to send the message that you're someone important who is worthy of such spending of resources. Once it becomes available to everyone, it no longer has that value.
Those wedding photos sounds like they are all about the bragging rights. And most likely actually going there, will continue to have much higher bragging rights. Might even get even higher if (AI) simulations become widespread.
Yeah I'll also use it. I agree that headshot imaging is a good use for AI, and other portrait styles too. For situations like weddings I take the point that a normal photo is what's required, but it could be cool to also have some fantasy style pictures - my nephew is getting married next year with a Viking themed wedding!
It cost $10,
according to the beginning of the article, not thousands, and it's not ego to be real and realize that a shitty pic that makes you look like an axe murderer is going to close off more opportunities than using a AI touched up profile pic.
How much is your time worth? Stable Diffusion is free, and the documentation to do it online is out there, but the time to learn it all to diy isn't trivial.
Personally I wouldn't use this, not because I have anything against AI, but because I think that trust is one of the main reason that you would use a headshot. If you have a headshot then people are more likely to believe that you're a real person, but if the headshot looks too perfect or too fake people start wondering if you're a a scammer using a stock image or headshot that's entirely AI.
Believability is going thru some growing pains. It's just as easy to gen normal bad photos as it is over-produced ones.
You don't have to use this but you'd be compared against people who use latest smartphones with awesome camera tech that enhances portrait pics automatically without any knowledge of where or ability to undo it.
It is quite striking that no two photos actually look like the same person - crudely similar, but the face shape is all over the place.
I would think this is an especially long “last mile” for training and fine-tuning: this might need a more sophisticated approach than a diffusion model, e.g an initial 2D->3D layer that first creates a 3D representation of the user’s face from a photo.
The vast majority of people don’t get professional headshots. They’re expensive and I’ve only ever had one when my employer has had a photographer come in for them.
They’re also a signal only for people you don’t know, not ones you do. I don’t know this guy but for some of these images, I couldn’t tell by looking at them that they are AI generated, and for the ones that are, I doubt a lot that when you have a tiny thumbnail of it on LinkedIn that you would be able to tell.
I think like many things, if they’re good enough but not as good as the best, it’ll still be enough for most people in the market for this.
Would have loved a comparison to some others. I haven't tried personally but been watching Pieter Levels post progress on fixing stuff like this for photoai.com and I think he has done a lot more tweaking than the service mentioned in the article which I hadn't heard of.
Most important part of the article, in my opinion:
> It’s worth pointing out that the company that owns the foundation model used to make these headshots, Stable Diffusion, is being sued for stealing over 12 million copyrighted photos to train it with.
I have to question how much these AI tools are profiting off stolen data.
I wonder why all services insist in going all-AI when we could instead go like 80%-AI and have the best of both worlds. Most people won't care about whether the background is not real, and some of the clothes were pretty good (except maybe the buttons). So why not generate a bunch of unimportant stuff and then add the actual person's face to the image? Maybe even the arms and hands? I think a much more useful service would be to take an amateur picture that you took from yourself (mobile phones' cameras are pretty good anyway in terms of quality and resolution) and then "enhance" it by adding a bunch of fake things like clothes and background.
It's because the people who dive headfirst into AI aren't interested in providing a good service. If you're lucky they are interested in using AI, if you're unlucky they are interested in getting paid for doing no work.
Since last week specialty smartphones and flux fine-tunes do "profession portraits".
Who really cares if people who want fake photos of themselves get them by AI, photoshop or some lens lighting makeup trick. They are for those seeking fame.
They shouldn't be worried just like programmers shouldn't be worried as these things are often subtly wrong in uncanny ways. People who know me immediately see it's fake. But most people don't know me, so it's good enough for many things and complete crap for others.
Often the discussion around these kinds of tech/AI dwell on the wrong things: questions like "will GenAI put artists out of work?" or "should photographers worry about AI-generated headshots?" miss the point that headshot photography probably only accounts for work for a small subset of working photographers, or that most artists are not making the same things that people use GenAI for.
Did Google Translate put professional translators out work? Probably not so much. Certainly not based on the discussions I've had with professionals, who may use AI translation but spend more time on copy-editing the results now.
Most of the time the AI-generated outputs are fine when you don't value the results more than the cost of getting better results. If you're doing translation of legal or technical documents, getting more accurate results is worth paying for.
Same thing for headshots: if they're important you probably pay for them, if not you weren't going to pay for a professional to do them anyway.
More than though: the whole point of portrait photography is displaying the individual, in all their idiosyncratic differences that make them distinguishable from others. These headshots have a blandifying effect of making the person look less like themselves the individual and more like a smoothed out, generic version of themselves.
27 comments
[ 0.56 ms ] story [ 1170 ms ] threadFor example, wedding photography. A picture that looks like you and your husband dressed up doesn't cut it. It's the moment, capturing everyone together, the crowds, etc.
The best I can imagine is that you could have someone walking around with a basic phone (or CCTV) wide angle capturing everything in low res, and then an AI could take those images and create convincing looking upscaled versions of them with interesting perspectives and composition etc. I still think though that people will want the real deal.
Sports photography and journalism are similar.
Another case, it's very popular in Asian cultures to go on a trip around Europe and get "wedding" photos in wedding dress in front of famous castles. There's no reason why these photos couldn't be generated with AI since the moment is fake already - and it's all about the photo. I worked for a company that was driving these people around, they spent max. 10 minutes at each location.
Ad the second case - we will see, but it definitely won't happen overnight, and there will be a lucrative market in between.
Even the article contains pictures with "weird hands"...the one where his arms are crossed in front the hands are WRONG :-)
even in all caps, calling it wrong seems to be underselling it. those hands are going to haunt my nightmares.
You don't have to use this but you'd be compared against people who use latest smartphones with awesome camera tech that enhances portrait pics automatically without any knowledge of where or ability to undo it.
I would think this is an especially long “last mile” for training and fine-tuning: this might need a more sophisticated approach than a diffusion model, e.g an initial 2D->3D layer that first creates a 3D representation of the user’s face from a photo.
The vast majority of people don’t get professional headshots. They’re expensive and I’ve only ever had one when my employer has had a photographer come in for them.
They’re also a signal only for people you don’t know, not ones you do. I don’t know this guy but for some of these images, I couldn’t tell by looking at them that they are AI generated, and for the ones that are, I doubt a lot that when you have a tiny thumbnail of it on LinkedIn that you would be able to tell.
I think like many things, if they’re good enough but not as good as the best, it’ll still be enough for most people in the market for this.
> It’s worth pointing out that the company that owns the foundation model used to make these headshots, Stable Diffusion, is being sued for stealing over 12 million copyrighted photos to train it with.
I have to question how much these AI tools are profiting off stolen data.
Who really cares if people who want fake photos of themselves get them by AI, photoshop or some lens lighting makeup trick. They are for those seeking fame.
Did Google Translate put professional translators out work? Probably not so much. Certainly not based on the discussions I've had with professionals, who may use AI translation but spend more time on copy-editing the results now.
Most of the time the AI-generated outputs are fine when you don't value the results more than the cost of getting better results. If you're doing translation of legal or technical documents, getting more accurate results is worth paying for.
Same thing for headshots: if they're important you probably pay for them, if not you weren't going to pay for a professional to do them anyway.
More than though: the whole point of portrait photography is displaying the individual, in all their idiosyncratic differences that make them distinguishable from others. These headshots have a blandifying effect of making the person look less like themselves the individual and more like a smoothed out, generic version of themselves.