Yes, but the problem is that the tiny sensors–while they have gotten more sensitive–are still a noisy mess most of the time. But with enough processing, that noisy mess can be made to look nice. The image processor is making guesses about what it sees and replacing the output with its best prediction instead of actual reality.
So far, these guesses have been fairly limited in nature, like enhancing faces or punching up the color or the like. We’re starting to see exceptions, like the moon thing, and we’re only going to go further down that road.
Even well-presented food tends not to look that great in photos without some work. What’s appetizing in person is not necessarily so in a static photo since there are so many other factors. A half-eaten plate of food looks just as appetizing as a full one when you’re eating it, but looks revolting out of context and in photographic form. An asymmetrical mound of mashed potatoes or whatever looks fine in person, but for a photo you want it to look perfectly formed.
Fair enough. The one person I know involved in this did both jobs. I don’t think stylists are necessarily toast either, but there will be a contraction in the market. McDonalds will still want their Egg McMuffins to look just so. But smaller local jobs are going to be reduced. On the other hand, all the food trucks that stop by me with fairly nasty-looking photos of their food will get a lot better. The AI will fix the color and lighting, clearly, but it’ll also put the food on nicer plates and remove the ugly background and tweak the food to be just a little more neat and symmetrical.
They are more than good enough to have done the job years ago. Light is light. And a decent camera phone with window light will do more than a good enough job for what many places need to do.
The moon thing is utterly ridiculous. Its like: we don’t even know what the camera is doing, if it’s enhancing existing detail, or inventing a new image. Which, again, for places that don’t use food photographers any more, may be advantageous. But it will have little impact on commercial clients who need a very specific image to match a very specific vision.
I was (among other things) a food photographer for over 10 years. I know what food looks like that has been presented “without some work.”
And Chefs create food that is pleasing to the eye. That’s what they do! The starting point is that the food will almost always look great. They certainly won’t plate anything they would consider “revolting.” You will already have a great starting point.
Yes, an asymmetrical mound of mashed potatoes will look like an asymmetrical mound of mashed potatoes if you don’t do any styling to it. But sometimes, that’s the point. Those badly taken photos of food you sometimes see at takeaways or diners often have me salivating. It takes me back to a certain point in time, to food that is yummy, and wholesome, and real. They don’t need better photographs. They don’t need to set unrealistic expectations.
This is exactly the market segment I was talking about. They already aren’t using food photographers. So AI isn’t going to have an impact on food photography here.
So why do they? I mean, there’s definitely some middle ground between the top and bottom images here:
They didn’t have to place the cheese so that the corners are perfectly symmetrical and droop down the sides just so, or that the egg would be a perfect cylinder, or that the meat would be exactly even like it is. The bottom image looks pretty bad, but if someone had spent an extra 30 seconds on assembling it it wouldn’t be so deformed looking. But the top image is what we see on TV and advertisements and their website, for obvious reasons. Making the food look perfect rather than just pretty good, but with the obvious flaws that you get in a normal environment makes a difference.
Yes, I agree. But there is a segment just a little above this–businesses that aren’t McDonalds and don’t demand absolute perfection, but want something with at least some of the usual tricks to improving the presentation–that might be taking business away from the stylists, if not the photographers.
…“what we got” strongly implies thats a photo the customer took, not the restaurant. Which kinda matters here, wouldn’t you agree?
The entire point of the bottom image was to show the food in a bad light. I wouldn’t be surprised if they actually spent an extra 30 seconds making it look worse for the photo.
Sure. And that’s the market segment I used to work with. I had a number of tricks up my sleeve that weren’t the ones oft quoted about stylists, but mainly about lighting, angles and composition, that made the food look great with the minimal of fuss. “Punching” the image up with a bit of AI wouldn’t add anything except to my bottom line, because I’d charge more for it.
Yes, obviously. That’s a pretty accurate representation of what the food actually looks like. They didn’t make it worse for the photo; that’s what Egg McMuffins look like when you unwrap them. I think that might actually be a better than average instance. The buns are actually close to centered instead of leaning off to one side.
If it weren’t McDonalds, but just some local place, the chef/cook might have spent a little extra time on presentation and taken the picture before it got smooshed up in the wrapping, and maybe put it on a plate. Then taken a picture with their cell phone. It would look a lot better than the bottom picture, but not as good as the top.
An AI trained on zillions of professional food pictures will have all those cues built in that can make an ok snapshot into something approaching the professional McDonalds picture. Not quite at the same level–but when everyone’s phone detects that you took a picture of food, and gives you a few options to punch things up a bit (more than just fixing the lighting), some places are going to opt for the phone instead of paying a stylist/photographer $1000 or whatever to get those improved shots.
…not really. That particular food item would look much different straight-out-of-the-kitchen. Wrapping it makes a difference. Putting it at the bottom of a bag makes a difference. Waiting ten minutes for a delivery driver makes a difference. Sitting in the passenger seat of the Uber drivers car makes a difference.
A chef typically wouldn’t be taking a photo of food in this state. This isn’t an example of what we were talking about.
And for most places, especially not franchise fast-food places with millions of dollars of turnover, it doesn’t need to be.
You mean trained on zillions of photographs taken by professional photographers, including me, that were used without our permission, nor offered compensation. An AI that if it weren’t for our hard work, skill, expertise and uncompensated labour, wouldn’t be able to do what it does.
Let’s just be crystal clear what we are talking about here. The only reason AI is any good at all is that us food photographers and food stylists and the chefs are great at what we do. We’ve invested thousands of dollars in gear, hundreds of hours in training. That food photo wouldn’t have that cool shadow wrapping around those mashed potatoes if the photographer had chosen an umbrella instead of a softbox, if they didn’t have a firm understanding of the inverse-square law. The AI didn’t need to learn or practice that law. It just had to “train” on zillions of photos.
AI would be utterly useless without us. And if it ever did kill food photography (and it won’t) then the pool of training data will get increasingly diluted with artificially produced facsimile’s.
And we are back to where we started. As I said from the beginning: there is a market segment that aren’t hiring photographers because they are doing it themselves. And better AI simply means that this market segment will be able to take better pictures. It may eventually have an impact on the larger market. But photographers will adapt, add AI if it works for them to their work flow, and charge accordingly.
There isn’t any game changing technology here. The examples of tech showcased here possibly offer a marginal, cosmetic improvement to a market segment that has largely shown it doesn’t really care about marginal cosmetic improvements. There is no money being left on the table here.
It’s a little of column A and a little of column B. The HW in camera modules has improved a ton since the first smartphones – pixel size, lens design, stabilization, HW HDR support, etc. Lately some of phone ODMs have been putting 1" sensors on their flagships. It has gotten insane.
But so has the image processing which has been transitioning to AI for the last few years – particularly on high-end devices. The bokeh effect, the de-noising, illuminant detection, color enhancement, and dynamic range.
All of these factor into a photo of a burger in a kitchen. Even more so if you add in low-light or faces.
This is very different than the generative AI that you two are contemplating, but there is a ton of camera AI already in the phone.
In 2010 as a joke I pitched a camera-less phone. It would use GPS to find where you are and download the best shot of that location from Flickr. You could be in front of the Tower of Pisa on a cloudy day and you’d get a nice, sunny day photo to show your friends back home.
But I think we are moving to toward this eventuality. With generative AI you can approach the camera sensor as a 3D prompt that tells the AI to generate the image you really want. As someone who is more of a sharpaholic than a noiseaphobe this is a disturbing idea.
However there is another perspective. Your memory of scene is a composite of ‘images’ from your eyes as well as other senses. It’s not stored in your brain as a .jpg. The highest-end camera doesn’t necessarily capture your reality. It’s part of why we struggle with image quality. What people want the image to look like isn’t necessarily what is there.
There’s a huge gray area between the two. On one end of the spectrum, things like smart denoising–not really generative, but still likely using neural nets that have been trained on real photos. In the middle, cases like the synthetic moon example I gave. They didn’t just paste a fake moon image on top of the captured one. They trained their AI on sharp astrophotography and their neural net figured out that a certain kind of blurry blob really looks like something else (a sharp picture of the moon). The moon pretty much always looks the same (given the same phase, etc.), so it probably didn’t take much for their AI to incorporate it.
And farther along the spectrum, not so far off, we’re going to see wholesale image transformation. You can already achieve this using online generation tools, so it’s just a matter of incorporating these into camera apps, and maybe toning down the generative aspect a little bit, so that you get something that looks like what you pictured, but better.
Camera apps are already doing some of this stuff, like cropping out parts that you don’t want, and having it generate the rest. But it’ll all get more pervasive, and probably default to on.
Yeah, absolutely. It’s not going to be quite “camera-less”, but the camera will just be one of several inputs to the image, and not the largest one.
Not just your memory, but what you see in real-time. We perceive far more than what our eyes capture. Our eyes have pretty shitty sensitivity and poor resolution outside of the fovea. Our eyes are constantly “painting” the scene and integrating the light collection over time. And taking our knowledge into account. When we read words, we aren’t somehow analyzing each squiggle independently; we just know that they’re text and can instantly decode them. Same goes for everything else we see. It all gets tagged against our memory, kinda like the robot-vision scenes from Terminator, but without the text–we just know that stuff intuitively, even though what’s going through our optic nerve contains none of that.
I’m apparently still not making my point clear. The top picture is virtually perfect in its construction and photographic elements. The bottom is what it actually looks like straight from the wrapper.
Someone at McDonalds could have spent 30 seconds just picking out a more symmetrical bun and assembling it more carefully. Then taken a picture with their cellphone. It would look perfectly adequate, just like your hypothetical chef snapping a picture of a nice meal.
But that’s not good enough for McDonalds, because there’s a difference between adequate and excellent. Generative AI will be able to take those adequate snapshots and make them better. Probably still not to McDonalds standards. But maybe comparable to a professional food stylist working the local scene.
Everything is built on what came before.
Some will, some won’t. There’s is a cost spectrum between being unwilling to pay a single dollar vs. McDonalds marketing budget. And there is a performance spectrum between cell phone snapshot vs. professional food stylist/photographer spending hours/days on a shoot. There is a certain relationship between the two based on the skills and compensation of human operators. Generative AI pushes the performance up for a given price. That will price some people out of the segment. Some will have to take a cut in compensation; others will have to increase their performance for a given price. Probably few people at the top have to worry. It’s those who are already in a marginal position that will get priced out.
Yet it doesn’t somehow look appetizing to me (as a foodie and a photographer). It looks sterile as fuck. The bottom isn’t that much better, either, to be fair.
It’s an Egg McMuffin, man. There are limits . Also, it’s fast food. The sterility–or rather the perfect consistency–is the point. Never mind that it’s not visually consistent in practice; that’s just a proxy for the consistent taste.
Also, that’s just one of several images I grabbed from @Stranger_On_A_Train 's link above:
It’s all the same basic idea. The real-life versions are just sad-looking. The Crispy Chicken BLT wouldn’t have been so bad if the person could have just folded a damn tortilla properly.
…the first picture is a bog-standard catalogue shot. Even lighting. Cut-out-in-photoshop. A headshot photographer could take that one. The skill here is being able to deliver images like this consistently, every single time. Something that oddly enough, AI actually struggles with. You don’t quite know what you are going to get.
Yeah, still not getting your point. You keep repeating everything I said. There are some people that don’t hire food photographers. AI will mean the people who don’t hire food photographers will be able to take potentially arguably marginally better photos.
It would still be useless without us. It isn’t building on anything we’ve done. The tools in this thread are gimmicks. Take our images (that were used without our permission) out of the dataset and it couldn’t do the job.
I’ve literally just explained all of this to you. Why are you taking the things I’ve just finished explaining to you, recontextualising them, and repeating them back to me?
I’m well aware of the scope of the market here. This is how I made a living for a number of years. I examined the market quite closely, both here and overseas. I wrote a marketing plan after a solid analysis of the market. I follow the trends, I keep up with my friends who are still shooting food in the industry.
Later this year I’ll be shooting a series of photography tutorials for YouTube. I’ll be including some videos on food photography. I can’t say I’m an expert in a lot of things. But I’m actually an expert here.
So what else do you want to know?
I just went back to check out Lunchbox, the “first AI Photo Generator” that was talked about earlier in the thread.
And its gone.
It was an offshoot of Lunch-boxes primary product, an enterprise-level ordering-system for restaurants.
This doesn’t surprise me in the least. I fully expect them, or somebody else, to try again. But there is a fundamental problem here:
There was no target market for this.
It wasn’t “punching up” food photos with things only a stylist could add.
It was generating fake-food based on the parameters of the prompt.
Who on earth needs that? What restaurant need random photos of fake food?
One of the use-cases mentioned in the advertising was for "common roadblocks like “coming soon photos.” But this isn’t a common roadblock. It isn’t a roadblock at all. It was an idea searching for a customer. Its the sort of thing that free stock photography websites like Unsplash and Pexels slot right in.
And the thing is, and I can’t stress this enough, the amount of “punch up” you would get from AI would be marginal to the point of nobody noticing. As you say yourself:
If a restaurant isn’t going to pay for a photographer, then they aren’t going to pay more than a few dollars for an AI food punch-up app. And if the app is needing millions of seed money and costs thousands every week to power the AI, how long will the company last if they are selling the app for maybe $10.00 on Google Play?
There needs to be a viable idea first before we can start to predict the downfall of an industry.
This is my usual thought about this sort of use case. If I’m going to have AI generate a fake cheeseburger or pile of fried chicken, why not just use a stock photo of a real cheeseburger? Unless my food product is so unique, I can almost certainly find a photo someone else did the work on. And if I am selling some frankly unusual menu item, I’m probably better off photographing it than trying to beat an AI into getting my unusual food item correct.
Two answers: first, AI-generated pictures are free; stock photography may or may not be. And second, I can get exactly what I want in the image. Even if the menu item isn’t unique, I may have something specific in mind. If I want it to show four fried chickens and a coke, I can get that. Maybe not available in stock photos.
But that’s not what I’m talking about anyway (I don’t know about anyone else). I’m talking about taking a photo of your own menu item and enhancing it. We’re used to simple enhancements like improving the lighting. And we’ve gotten somewhat more generative tools like being able to crop out something unwanted and using infill to fill the gap.
And this is just going to be taken to the logical extreme. I take a picture of my own fried chicken, but maybe it doesn’t look so great. There’s junk in the background. The basket has some grease stains. The chicken itself just doesn’t look as perfectly crispy as you’d want. The fries are kinda limp. Whatever–but the AI tool has been trained on zillions of professional photos where all this stuff has been made absolutely perfect. And so you’ll take a picture, it’ll give you some enhancement options, and bang, you have something that looks almost professional. It wasn’t generated completely from scratch, but what you’re seeing is basically a fake.
Phone cameras pretty much already do this for portraits. Put it in portrait mode, and it gives you some nice bokeh in the background, smooths out your skin tone, eliminates some blemishes, maybe even tweaks your facial proportions a tad. Can look a bit fake at times, but generally seen as an improvement. Faces are easy for AI these days, and a popular application, so that’s one of the first things to be worked on. We should expect the same for food and architecture and all sorts of other things.
I own a local restaurant (hypothetically). I take my own pictures for the menu, just to illustrate the items. But I also pay a local guy $1000 to take some glamour shots for the menu cover and the website. I want these to look good, but I also can’t afford much.
The new AI tools can take one of my marginal snapshots and make it look professional. It’s not inventing something from whole cloth; it’s taking my photo and enhancing it. But it’s a step beyond just fixing the lighting–it’s adding details that weren’t there before, just like Samsung’s camera app look a blurry picture of the moon and added detail from other photos it was trained on.
We aren’t quite there yet, but it’s such an obvious next step, which is already in process, that it’s basically inevitable. And that restaurateur will use that instead of paying the local guy.
No one. Not what I’m suggesting, and it sounds like a dumb idea.
The vast majority of these tiny AI startups will fail. They have no real business model, and mostly know it, and are mostly just hoping to get bought out by either a company trying to jump on the AI hype train or who just want their team.
No, I think the risk is from Apple and Google (and to a lesser extent Samsung). They want to sell their phones. They are constantly adding new AI features to their camera apps. And have millions of customers that love posting food pictures to Instagram or wherever. These companies have vast resources and can optimize for a tiny percentage of their userbase.
No one is paying for these things directly; Apple and Google do them because they think it’ll attract 1% more users to their platform. And because they have buildings full of AI training hardware already.
I don’t think the food photography industry will collapse in the near future; just that it’ll get harder, and require fewer people. Used to be that every mall had a portrait studio. Now, they don’t. Portrait studios still exist, but they’re kinda rare.
I was in the first run of people excited about colored smudges in Nightcafe that maybe look like a thing if you poke yourself in the eye and squint. I’ve had a MidJourney sub for over 18 months. I have about 40 different SD1.5 and SDXL models on my PC, plus have trained my own models and LoRA. I’m aware of what AI imagery can do.
I would still look for a stock photo before settling on an AI image because the AI image is either likely going to be “weird” in some way (and those odds go up the more specific you want your picture) and because the time spent fixing a weird photo may easily exceed the time it would have taken to trawl “Fried chicken” photos on Pexels and the like.
I think Banquet_Bear’s point is that the area of food photography where AI might assist is already eaten up by existing cell phone tech. At the low end, no one cares what taco clip art you slap on your sunfaded menu board at your roadside stand. On the high end, someone putting an ad in Fancy Dinin’ magazine or national ad campaign is likely going to want human staging and eye to attention. And, increasingly, people who want a photo of their bakery case for Instagram or a pizza photo for the Valpak mailer can accomplish that with an iPhone. And can move their own salt shaker and take infinite reshoots rather than relying on AI to do the heavy lifting. If there’s grease stains on the basket or the fries look limp, you just grab a different basket and rearrange some new fries and press the phone’s button fifty more times, you don’t call a food photographer.
In other words, half of the stuff you mention, the phones can do now (lighting, photo effects) and the other half probably isn’t (in Banquet_Bear’s opinion) going to make any immediate and significant difference with the sort of people who need professional food photography.
. Also, it’s fast food. The sterility–or rather the perfect consistency–is the point. Never mind that it’s not visually consistent in practice; that’s just a proxy for the consistent taste.