We are far from it . It could make for a great script for a Black Mirror episode do 

This could arrive but I do think micro is quite safe against this type of image creation because the higher costs and time needed for creating the final image. But in much more expensive projects it is already happening.

I have worked some years in the advertising sector and 3D is used extensively specially in car advertising. But the complexity is big and many times it is much cheaper to move a production crew across countries to shoot real scenes and maybe add later some elements in post that do it all from scratch. For simpler backgrounds (a desert, mountain view, etc)CGI can replicate reality very well but with more complex backgrounds the time involved in the creation makes it more easy to shoot the real thing.

In micro or medium stock (even macro with nowadays low prices) making a 2-3 hours search and buying is a thousand times cheaper than going a CGI route.

I think that mathematics is in our side and someday computers with AI will make this slow process so much faster and convenient that it might be an option for stock users. But I doubt our generation will see it. With the present computing speed bottleneck, Moores Law is no more a reference and it might be well a lot of time to have the computer power to render CGI properly and at the required speeds that stock buyers need.

AI can make decisions but it still requires technology to execute things.
With today imaging technology (Photoshop) it is impossible to do realistic composites. My opinion.

My answer is no, this will not work.

I was a software engineer for about 25 years.  I worked on everything you could think of.

Relax.  There isn't any "Artificial Intelligence" - it's 99% marketing hype.  Yes, software is doing some things it couldn't in the past - like recognize faces. But it's all just algorithms and data; newer, better algorithms, more data, faster processors.  Yes someone could write code to recognize emotions in facial expressions, or gender, or beauty.   And the results would be comical.   We are still light years away from software that could recognize a "good photo".   

I think that most important point is not the "AI" software efficiency.

The point is that a buyer should know EXACTLY what he wants. And for conceptual images I'm quite sure this is not real. The buyer simply doesn't know exactly which kind of image want.
This side of human creativity will never be overtaken by any artificial AI.

Well, I hope so :-)

Why go to the trouble, if you need bespoke images you could have remote photogragphy at your finger tips.  Mini airports all over the world stacked full of drones accessible from across the planet, we have the technology. Travel photography without the travel.

Quote from: stockastic on April 13, 2017, 12:49

I was a software engineer for about 25 years.  I worked on everything you could think of.

Relax.  There isn't any "Artificial Intelligence" - it's 99% marketing hype.  Yes, software is doing some things it couldn't in the past - like recognize faces. But it's all just algorithms and data;

Actually it's not. I too have worked in software development for nearly 30 years and was involved in AI through the so-called 'AI winter' where the computing power was not sufficient for the challenges or the hype.  In the first decade of this century computer vision was about algorithms. In this decade it has moved away from algorithms such as SIFT to machine learning (deep convulutional neural networks) which are far more effective. The output of an algorithm is predictable, with machine learning it is not. The GoogLeNet winning entry in the 2014 ILSVRC competition for visual recognition used a D-CNN (with a week's 'training') and achieved 93.3% accuracy.

I suggested the replacement of stock photography by automated image creation on this forum about a year ago.  Now I think a practical application is about 5 years away though I doubt it will be worth implementing in the stock industry.  The first requirement of it is the image recognition part.  There will probably be enough images available to search that the image someone describes in detail will already exist and the cost of paying the creator (what will that be then? 3 cents?) will probably be less than the cost of the computing power to generate the image.  And, as buyers never know exactly what they want, a few pages of pics to choose from is likely going to be preferred in conjunction with a 'more images like this' option to a single image of 'this is what you asked for'.

The only sure prediction I can make is that whoever perfects the technology it won't be Shutterstock.  Implementing D-CNNs is quite a bit harder than keeping a database online and managing changes through development and test environments to a live site.

Eh.   Whether the code is procedural, declarative, or uses some less deterministic method to converge on a solution, it's still just software written by humans. There aren't any actual neurons in a "neural net", that's just a cool buzzword for a way to structure some code.   Any algorithm that uses a random number generator is less than "predictable".  If it iterates or recurses, while saving intermediate outcomes and using them to improve future guesses, I guess today it could be called "deep learning" and  the marketing guys will love it - but color me unimpressed.

I did some work for a biometrics company, I saw facial recognition technology,  there's no "intelligence" in it - just raster scanning and extracting shapes from gradients, normalizing them spacially, and hunting for a configuration that might be 2 eyes above a nose and a mouth.

IMHO,  AI is where it's always been - a distant dream.   Of course, facial and voice recognition, and language processing, have progressed well.  Marketing has simply decided that "AI" now exists, because it sells like crazy and, well, it's in all the movies so people think it's real.   All the big financial companies are "using AI" - or better yet "using an AI" - instead of just crunching numbers.  It makes me trust them so much more :-)   I see the term "General Artificial Intelligence" or "GAI" has been coined, to describe systems that actually recognize and solve problems, work towards a goal, and do other things that seem closer to human thought processes.  Plain "AI" now seems to mean any program doing something new or impressively difficult.

Ok, putting my sarcasm and ignorance aside, I'd like to know more about what's going on in neural nets today; I'm out of it now, I quit a few years ago.  I see great progress, I'm not a Luddite, I just object to the huge overuse of the term "Artificial Intelligence".  No doubt, with enough computing power, code will be written that can scan a photo and extract "2 happy people, one male, one female, both over 65, with bicycles, near a wooded area".  But to get from that to anything I'd call AI is still a matter of light years.

OK, instead of the contentious term "AI", please instead read "powerful computers and advanced software" 

Well, turns out linear regression is a machine learning algorithm. Who would have thought 40+ years ago? We are in good hands...