Hacker News Reader

Comment by x__x | Hacker News Reader

Surprised there was no mention of the upcoming threat of AI stock photos, and how easy it will be to generate massive amounts of these fast and cheap. Ne real people, who won't have to sign contracts or be paid.

https://www.theverge.com/2019/9/20/20875362/100000-fake-ai-p...

View on HN

Lately I've been imagining what it would take to create an AI-generated food image. It's super hard!

Just the rights management issue of the training set is a tough question. Then you have to account for quality, direction and dimensionality of light, food styling, props, custom products or ingredients, the list of challenges goes on an on.

Then after the image is produced, if you're a food company, all your imagery is a kind of 'claim-based' imagery meaning consumers assume it is real and in some cases, there are laws preventing the use of fake foods in ads.

I'm curious about the implications of a world where AI stock photos are ubiquitous, but some styles of imagery are pretty far out at the moment.

Source: Spent 15 years in the commercial photo industry.

I do wonder if the rights management issue is impacted by the decision referenced in this article regarding training machine learning models on copyrighted material:

https://towardsdatascience.com/the-most-important-supreme-co...

https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,....

It seems like part of that case hinged on economic impact. If you're using copyrighted books and show snippets of the book that's one thing. But using copyrighted material to build something that is attempting to replace the copyrighted material seems like it could have gone the other way.

> Lately I've been imagining what it would take to create an AI-generated food image. It's super hard!

I can’t wait for a chef to throw their arms up because they get tasked to create an AI-generated non-existant food tray that performed well in Amazon Turk focus groups.

I would love a image generator that would make images based on the ingredients. When creating a recipe database, not only do you have to photograph thousands of dishes, you also have to cook them.

The (human) models and photographers will be in real trouble when generative models are able to make an ad on the fly with a fake person who looks like your neighbor with the best landscaping using the advertised lawnmower.

Um, last I checked, AI models of humans looked super, super creepy. Not the sort of content you would want on a marketing campaign, that's for sure.

When the AI worked it worked well but when it failed the failure was obvious and creepy.

Some looked creepy, others looked average, others looked practically real. You can still retouch them a little using Photoshop and you have something that is 99.9% real. Good enough for a campaign.

It's time to check again. The latest ones are indistinguishable from real humans, at least to my eye. Somehow the uncanny valley has been crossed already.

Form what I remember that AI could only do one pose.

This has changed. It's hard to tell they're fake. For example, each reload of https://thispersondoesnotexist.com/ produces a new computer-generated image of a person.

You're right, that's better than it used to be by orders of magnitude. However, there remain things that "just don't look right". The most obvious I can spot is the teeth, followed by warped edges (odd indentations, a missing ear that ought to merely be partially hidden), followed by the eyes(more subtle, but definitely there). The "big swathes of skin" are generally fine. There are also wrong numbers of teeth, partially missing eyebrows, etc. Finally, there's not much correlation between apparent ethnicity and facial characteristics or hair. All these "just look wrong" enough to put someone off of a marketing brochure.

These are random images though. If you do want something specific, you can dial the right knobs to fix the issues you mentioned and match up the features you don't like. Adding some extra constraints should be possible if you actually want commercial usage.

I reloaded the page ten times and every single one had something clearly off.

It's hard to generate perfect images at high resolution. The fact that they are even plausible at high resolution is very impressive. If these were downscaled to 128x128 you would literally not be able to distinguish real from fake.

You mean like the Goldeneye box art? He has two mouths and it didn't stop the game selling like hotcakes.

https://gaminghistory101.files.wordpress.com/2014/11/goldene...

Where’s the second one? I see one mouth and a fist (holding the gun) in front of it.

https://screenrant.com/worst-video-game-box-art-mistakes-obv...

> https://thispersondoesnotexist.com/

This is exactly why I read hacker news. This is both perturbing and amazing all in the same time with all the links you need to understand how it works under the hood. Thanks!

several of them I looked at had malformed ear (I assume the training set had earings and gauges and such mixed in) and one had a tuft of of hair growing out of their neck

When was the last time you checked? Because the last time I checked I found it hard to tell the difference.

Once AI is good enough to take photos better than a human, I think AI will be good enough to make entire categories of jobs obsolete. Photographers will at least have a fun hobby to enjoy while collecting their basic income.

Reply on news.ycombinator.com