Author Topic: Ethics of photo synthesis (Read 2936 times)

adijr · « **on:** February 06, 2017, 21:38 »

I'm curios what people think of photo synthesis algorithms, and how this will impact [stock] photography in the future.

Although still far from perfect, photo synthesis algorithms are starting to be able to generate photos in interesting scenarios (for example, given an input sentence: https://arxiv.org/abs/1612.03242). These algorithms are "trained" on huge collections of photos (e.g. imageNet). Depending on the algorithm, they will create completely novel images driven by some mathematical costs, or they will be semi-copying small parts from other images.

What do you think is the ethical (and legal?) situation for using images created by these algorithms. In some sense, they can generate completely novel images, but they are trained from existing photos. Better, yet, what happens if this algorithm is trained on photos from an artist that are under a non-commercial license, and then generates brand new photos in that artist's style?

I believe these algorithms are coming, and some very deep conversations need to be had

. If anyone has any links to other similar conversations, they would be appreciated.

Thanks.

niktol · « **Reply #1 on:** February 06, 2017, 22:29 »

I don't think people who are seriously working on computer vision and pattern recognition are interested in stock photography now. However, I am pretty sure that developments in machine learning will have a tremendous impact on many professions, not just a small group of photographers. Wait for the data-based approaches in diagnostics and medicine in general, to mention just one. Lotsa kids who go out of their way to get into med school will be seriously disappointed.

Jo Ann Snover · « **Reply #2 on:** February 06, 2017, 22:33 »

Sounds analogous to the issues of sampling in the music business.

https://www.nolo.com/legal-encyclopedia/permission-sampled-music-sample-clearance-30165.html

I don't suppose anyone will want to participate in detailed discussions until there's something pretty close to ready - do you think this is a year or two away? Five years?

I'm guessing that agencies won't take anything that you don't own the rights to. They won't be looking at ethics but managing their legal risk (agencies are risk averse).

Various sharing sites will use all sorts of copyrighted stuff as input and hope they don't get identified.

If it matters to the finished image which pieces are used as input - versus any pile of pixels with the right colors will do - it'll be harder for the creator to claim copyright in the new work. If enough people wanti to create this type of imagery, possibly agencies will have a new type of license for use of image chunks in new works to make the synthesized images legit. A new type of extended license.

niktol · « **Reply #3 on:** February 06, 2017, 22:46 »

At this point there is a chance that there won't be any agencies.

adijr · « **Reply #4 on:** February 06, 2017, 23:51 »

Quote from: niktol on February 06, 2017, 22:29

I don't think people who are seriously working on computer vision and pattern recognition are interested in stock photography now. However, I am pretty sure that developments in machine learning will have a tremendous impact on many professions, not just a small group of photographers. Wait for the data-based approaches in diagnostics and medicine in general, to mention just one. Lotsa kids who go out of their way to get into med school will be seriously disappointed.

Oh I absolutely agree.

But someone can take advantage of this (and not always necessarily in a malicious way). For example I think it raises an interesting (and legal?) point -- one could download Trey Ratcliff's entire New Zealand collection, generate new images, and sell (or use) them. Should Trey get some (credit/royalty) for creating the "training" dataset?

adijr · « **Reply #5 on:** February 06, 2017, 23:54 »

Quote from: Jo Ann Snover on February 06, 2017, 22:33

Sounds analogous to the issues of sampling in the music business.

https://www.nolo.com/legal-encyclopedia/permission-sampled-music-sample-clearance-30165.html

Nice analogy! But for some algorithms, it goes deeper/more subtle than that, i think. Say the algorithm doesn't copy any exact soundbit, but just a general "feel" of the song.

Quote from: Jo Ann Snover on February 06, 2017, 22:33

I don't suppose anyone will want to participate in detailed discussions until there's something pretty close to ready - do you think this is a year or two away? Five years?

I'm guessing that agencies won't take anything that you don't own the rights to. They won't be looking at ethics but managing their legal risk (agencies are risk averse).

Various sharing sites will use all sorts of copyrighted stuff as input and hope they don't get identified.

If it matters to the finished image which pieces are used as input - versus any pile of pixels with the right colors will do - it'll be harder for the creator to claim copyright in the new work. If enough people wanti to create this type of imagery, possibly agencies will have a new type of license for use of image chunks in new works to make the synthesized images legit. A new type of extended license.

If one trains one of these algorithms on someone's portfolio, but then they generate a completely fake image in the *style* of that artist, what *is* the legality of this? There's no explicit copying involved, and the subject matter might have changed completely, even. But without the portfolio as training, the generation would not be possible.

Justanotherphotographer · « **Reply #6 on:** February 07, 2017, 02:06 »

niktol · « **Reply #7 on:** February 07, 2017, 10:47 »

Quote from: adijr on February 06, 2017, 23:51

Oh I absolutely agree.

But someone can take advantage of this (and not always necessarily in a malicious way). For example I think it raises an interesting (and legal?) point -- one could download Trey Ratcliff's entire New Zealand collection, generate new images, and sell (or use) them. Should Trey get some (credit/royalty) for creating the "training" dataset?

It's happening right now without all that hassle. Artists are inspired (or "inspired", if you prefer) and imitate other people styles. Sometimes a new style is born, it's only natural. I don't think anything can be done about it, legal-wise. Artists are not locked up in a box from birth, they see what other people are doing. Again, a training set is nothing without an algorithm to it. I guess some IP goes with the novelty of the algorithm. And I can imagine that it will have nothing to do with just sampling other people visual elements, it will go deeper than that.

Basically, there will be programs that create images for you according to your liking. Agencies will go the same way as Blockbuster when it started streaming videos instead of renting VHS. Oh wait...