[thephyber]

While I agree with what you say, I think there's also subsets of modeling which can be cheaper with a computer which doesn't need wages, an agent, travel, or royalties. Also remember that models are human talent and human talent tends to come with costs which must be geographically local: makeup, photographer, scene, etc. Humans can only work so many hours a day, they can develop drug or eating habits, they can age. They can say indecorous things which will cause an outrage mob to want to boycott brands associated with them. All of these are costs or liabilities.

Granted, there may be a new generation of agents that specialize in AI models and royalties may still exist (with shrinking margins), but if nothing more, AI is likely to opt downward pressure on wages/jobs/contracts some of the non-minimum-wage models. Once it's bootstrapped, it will either become more appealing (for the reasons I mentioned above) or turn out to be complex and not worth the cost/risk. Only time and experimentation will tell.

[kbenson]

> Human talent ... can say indecorous things which will cause an outrage mob to want to boycott brands associated with them.

I'm not envisioning a future in which some company creates a full persona for their AI models, and we get a full Tay[1] moment out of it, and then we've come full circle.

1: https://en.wikipedia.org/wiki/Tay_(bot)

[dredmorbius]

There are also the companies producing or providing the AI. Any entity capable of producing a convincing AI fashion substitute can do much more besides, some possibly not wearing well with public opinion.

I'm reminded of the (aweful) film Simone (https://en.wikipedia.org/wiki/Simone_(2002_film)), and a Ben Bova novel (possibly Starcrossed) in which CGI gradually replaces human actors.

[kbenson]

To clarify, in my original post, not was supposed to be now, as in "I'm now envisioning"

Lots of typos for me today... :/

[dredmorbius]

Right, I realised that.

Am also agreeing with you and extending the notion. AI is by no-means outrage-proof.

[kbenson]

Sure, your comment didn't look like it was trying to disagree as much as add more info, whether you realized I had a typo or not. I had just happened to read a comment I had mangled atrociously earlier because I did it on my phone and then I saw this comment, and the urge to correct myself here was overwhelming. :)

[dredmorbius]

I feel that urge. Soft-keyboards, and HN's edit window, are treacherous.

[joe_the_user]

Also remember that models are human talent and human talent tends to come with costs which must be geographically local: makeup, photographer, scene, etc.

This sounds vaguely similar to initial discussions I remember with self-diving cars. It begins with "I look forward to the day we don't have fallible humans at the wheel" and end with the realization the AIs are going to be more fallible and in different, strange ways.

I mean, hypothetically, suppose you had a system that could constantly monitor world fashion trends, world clothing markets, the routes of hip and average people and so-forth. In the case you might create a system that produced a variety of still and moving images that satisfied all the constraints that today's fashion industry satisfies just with a few written or spoken suggestions from executives. Then you'd eliminated no just models but a large chunk of the industry.

But let's look at what "AI" is now (and what it seems likely to be for a while without any "revolutionary" changes). What you have now is a way to extrapolate typical objects out of a stream of similar objects. Just GPT-3 does a great job create texts that sound vaguely right, you can create a vaguely plausible looking set of still and moving images of one or another "typical" model. Moreover, these extrapolations require constant training by professional much more highly paid than actual models now (as the GP notes).

Further, without being in the fashion industry, I'm pretty sure there's a lot more to a useful set of images than "looking about right". I suspect you could generate a model-image that would "work" with a kind of clothing (since both clothing and image can be trained). But generating a model-image that suits a given demographic, that expresses "what's becoming hip right now" and so-forth would be extremely hard. It may not be impossible but it would require lots of high paid labor by AI engineers, defeating the entire purpose once the novelty wears off. And all this is to say that these "replace human activity" approaches wind-up with the problem of doing an "90%" of the activity right and then foundering on corner cases - like self-driving cars that are easy-yet-impossible AI tasks.

[derefr]

> I suspect you could generate a model-image that would "work" with a kind of clothing (since both clothing and image can be trained). But generating a model-image that suits a given demographic, that expresses "what's becoming hip right now" and so-forth would be extremely hard.

I mean, when I picture the problem, I imagine taking a single panorama-like photo capture with a smartphone, circling repeatedly around a human being whose likeness you want to use; and the phone using its barrage of sensors and ML cores to spit out a pre-rigged and textured high-poly 3D model of that person, that you can then drop into Blender and throw your clothing designs onto (i.e. the very same digitally-simulated designs that your designers prototyped with before getting the design made for real — presuming there was any amount of industrial design going into the object, which there certainly is for anything as complex as e.g. glasses frames, or a handbag.)

The pre-rigged 3D model output from such a body-scanning app would have a standardized rigging, such that 1. you'd know how your digital clothing items would interact with it before attaining the body-scan itself; and 2. allowing you to throw some posing "behaviour" scripts on it (that target said standardized rigging.) So this could all be parallelized.

Last step: pick a 3D-recreated environment, set the viewpoint camera and lighting, and snap screenshots at will. (This part doesn't need to be a science; you can just put a trained photographer in VR goggles, and have them circle around the digital model taking digital pictures with their field-of-view at time of trigger-press being the composed shot.)

The important part of this, from a cost perspective, is that you can then reuse this model for a combinatoric number of "shots", without ever paying the original body-scanned person again, or taking time to organize a new physical shoot with them. You can "re-shoot" them in localized advertisements for every target market you're launching the product in, all without needing to leave the room, let alone paying them to come back in. If you launch accessory products months later, you don't need to retain their talent; you have them "on file." Likewise if you need to dredge them up 10 years later for an anniversary "shoot."

[jameshart]

> But generating a model-image that suits a given demographic, that expresses "what's becoming hip right now" and so-forth would be extremely hard

Just generate a spread of different looks, test them on different audiences, and measure engagement. Feed that back into the parameterization for generating the next set of images.

[thephyber]

> And all this is to say that these "replace human activity" approaches wind-up with the problem of doing an "90%" of the activity right and then foundering on corner cases - like self-driving cars that are easy-yet-impossible AI tasks.

I think this is the crux of the issue.

[Talanes]

I don't think the boycott mob is a unique problem to the human. Could just as easily end up with an anti-AI boycott mob. Neither bet is fully safe on that front.

[thu2111]

The issue is for companies that just want to sell clothes, without also being forced into a wide variety of shallow, poorly thought out political stances that could blow up in the companies face. Note just how much of this article is the 'model' talking about what is basically politics. They're half model, half politician. This is 100% liability to most of their employers, who ultimately just need a good looking mannequin to wear some clothes or pose seductively for advertising.

A GAN generated face textured onto a convincing 3D model means you can just keep clicking or adjusting until you got a truly beautiful model, without needing to pay extra for it, and you know your exposure is capped in entirely predictable ways. This model won't unexpectedly cause trouble for you, or demand you take a particular stance you may not agree with in order to employ them.

[thephyber]

I don't see the average consumer caring whether the model for their perfume commercial is flesh and blood or CGI. They already don't care what the quality of life is like for the test animals for those same cosmetics. Plus, it's easy enough to create fake personal social media profiles to pretend like the persona has a real history.

[lazide]

The difference is, the bot can be made to only say what you want it to. That is a LOT harder to do with a human.

If you're a big outfit looking to control risks, it might be very tempting.