Adobe Photoshop Beta with Generative AI

[Manfred M]

I downloaded the beta Photoshop from Adobe a few days ago and started experimenting. I used the bicycle shop owner in Srinigar as my subject, made some selections and ran it through the Generative AI function.

The results ran between somewhat convincing to absolutely dreadful. Here are a few shots from my experimentation...


This is the original image:







I wondered what he would look like in a trench coat?








It was pretty chilly, so perhaps a puffy down-filled coat?







Then I went wild and put him in a tutu.






While it is early days for this technology; I've been asked to do a 5 - 10 minute talk on the subject at one of the photo clubs I belong to at the beginning of the next club year (September / October), so I have started my research...

[Cantab]

Manfred, I'm not sure what I think of this. At the very least it takes photoshopping to a new level. I assume that the process was relatively straight forward from your user perspective?

[Manfred M]

Manfred, I'm not sure what I think of this. At the very least it takes photoshopping to a new level. I assume that the process was relatively straight forward from your user perspective?

The scary part was how simple it was. I used the Quick Selection tool and selected his clothes and typed in "replace clothes with...."

In one case it was a trench coat, in the second a down-filled coat and in the last one a tutu. About one out of three attempts gave good results (I did no retouching on the output).

Sometimes the results were a bit bizarre, just like we see in other forms of generative AI image creation. When I tried to put him in green medical scrubs, it included a bowl around one leg...

[Manfred M]

Another example is where I used the Generative AI as a "smart" sky replacement. In the standard Photoshop / Lightroom algorithms, just the sky is replaced and that can leave all kinds of artifacts. In this image I did a sky selection and directed the software to soften the sky. It gave me three different versions; I picked one and did a bit more burning down around the areas where the sky and the grasses meet.

The first image is the original and the second one is what came out of the Generative AI, plus a bit of retouching.


Original Image:








Generative AI = use of Photoshop Burn tool

[pschlute]

Tutu wins for me

[Dave Humphries]

Interesting to see Manfred, thanks for posting - they seem to confirm what I'd seen (in a YouTube video) regarding AI's ability to reproduce hands and fingers though.

[pnodrog]

Tutu wins for me

I'll have to revise my impression of you...I wonder if it's possible to ban you from getting a copy of Photoshop AI?...

[billtils]

I don't see any value in having this sort of thing included in a top level photo editing program. To me this is pointless and disapointing compared to what the same company has done with its AI noise reduction routine in LR (full disclosure: the last comment is based on what I've seen here and not on direct experience).

[BobGilbody]

I take it you first obtained permission from the gentleman prior to displaying him to the world as a cross dressing ballerina.

[Manfred M]

I don't see any value in having this sort of thing included in a top level photo editing program. To me this is pointless and disapointing compared to what the same company has done with its AI noise reduction routine in LR (full disclosure: the last comment is based on what I've seen here and not on direct experience).

First of all, it is important to remember that this is a beta test product and not one that has been introduced as a mainstream product. I do see potential in what it will be able to achieve in the future; much like some of the other enhancements we have seen in the past. I love the Quick Selection tool but it doesn't always to a perfect job, so I end up tweaking the results. Same comment goes for the up-sampling tool and I have to spend time cleaning up the artifacts. The Content Aware products can introduce some strange artifacts as well, so knowing their limitations makes my job easier and possible.

Like anything else, these tools speed up what I can do. If I wanted to, I could take a model into the studio, set them up in a trench coat, down-filled coat, tutu or scrubs and light them to match the scene and produce results that are much stronger than what I get out of the current Adobe Firefly product. Something like that would take a number of days to pull of, rather than a minute or so, so it is a step in the right direction.

The last example where I used this technology for a background / sky replacement is where I would start with. It works much better than the current Sky Replacement software.

I suspect things will be a bit more polished when (and if) it is productionalized and comes out of Beta.

[Manfred M]

I take it you first obtained permission from the gentleman prior to displaying him to the world as a cross dressing ballerina.

No and legally I do not have to... He certainly knew I was taking his picture.

I own the copyright on original image and I can do whatever I want with it. Ethically, I thought about it for some time, but as this is largely use for educational and evaluation purposes, I don't really have any qualms here. I'm not making any money off the image nor am I giving anyone the right to reproduce and share it.

That being said, I have no issues taking that particular version down, if requested.

[billtils]

No problems with anything in the software that increases efficiency but nothing you say here changes my view on the "coat replacement" function. Using it to road test an idea and then as a guide on how to set up and photograph a scene is fine, but using it to create the final product is altogther different and looks to be a line that would be too easy to cross.

[DanK]

I largely agree with Bill.

There's AI, and then there is AI. It doesn't make sense to me to drop into the same bin everything that has been produced by machine learning. Street drugs and legitimate drugs rely on the same chemistry, but we don't treat them as similar in effect or desirability.

I have generally have no problems with machine-learning-based techniques that simply allow one to perform an action more efficiently. I use Adobe's subject select frequently, for example. It makes editing candids of kids far faster. Based on my experience so far, I expect I'll use the AI-based noise reduction often as well.

However, this generative fill is inserting someone else's content, modified by the software to fit the specific shape you have to fill. Those clothes presumably came from the training set, which is other people's images. The software had crossed that line some time ago. Sky replacement using one's own images of skies is analogous to subject selection, IMHO. It's also analogous to a painter who decides to paint a sky differently than they see. But substituting a sky from the set provided by the software is using part of someone else's image. It's like directly copying another painter's sky.

I posted before a comment that I think bears repeating. A friend of mine who spent his career in tech and was dabbling in travel photography said numerous times that what would serve many people better than a real camera is a device that reads location via GPS and presents the "photographer" with a library of images take from that spot that are better than what they can do themselves. That's pretty much where we stand now: we can "create" images that are increasingly not our own.

That's a path we actually started on without AI. Programs like Nik allow people to drop on styles that someone else created. You can just page through them and decide after the fact what you like.

There are plenty of aspects of modern life where that's true and where it doesn't bother me. I'm old enough that I used to tune my own car and do routine maintenance. Timing was simple to describe and not that hard to set. Now I can't do anything at all to maintain my current car, a VW ID.4, and in fact, I often don't actually know what it's doing. I frequently find out that there are additional functions going on under the hood that I didn't know where there. That's fine: I am not spending my free time as a car designer or mechanic. If it's reliable, safe, and enjoyable to drive, I'm a happy camper. It's the result I'm interested in. In contrast, I want the satisfaction of creating my own images, and this sort of software will erode that.

[billtils]

+1 to what Dan said, and an extra "+" for expressing it so well.

[BobGilbody]

No and legally I do not have to... He certainly knew I was taking his picture.

I own the copyright on original image and I can do whatever I want with it. Ethically, I thought about it for some time, but as this is largely use for educational and evaluation purposes, I don't really have any qualms here. I'm not making any money off the image nor am I giving anyone the right to reproduce and share it.

That being said, I have no issues taking that particular version down, if requested.

"I own the copyright on original image and I can do whatever I want with it."

Sorry didn't know that.

[DanK]

And by chance, I stumbled on this after posting. A more complicated version of exactly what my techie friend suggested years ago: an AI-powered device that makes it unnecessary to take a photo at all:

https://bgr.com/science/crazy-new-ai...-does-it-work/

[Cantab]

I take it you first obtained permission from the gentleman prior to displaying him to the world as a cross dressing ballerina.

No and legally I do not have to... He certainly knew I was taking his picture.

I own the copyright on original image and I can do whatever I want with it. Ethically, I thought about it for some time, but as this is largely use for educational and evaluation purposes, I don't really have any qualms here. I'm not making any money off the image nor am I giving anyone the right to reproduce and share it.

That being said, I have no issues taking that particular version down, if requested.

Manfred, I'm not sure that the copyright issue leads to the legal conclusion that you can do whatever you like with the image. There is still a question of informed consent by the subject in the image.

Although I retired from legal practice a number of years ago, your comment set me wondering how I'd have responded if the man in your photo had arrived at my office, with a copy of the images in this thread. He might well have told me "Yes, I consented that this unknown photographer take a photo of me. But I would never have consented if I'd known what he might do with the photo of me: place on the internet a photoshopped image of me wearing a tutu. People that I know have seen this image and laughed at me."

There are clearly situations where the holder of a consensual photo is not at liberty to do certain things with the photo. I don't know whether the images in this thread cross the legal line around informed consent. But I do not see the issue as free from doubt. As you've noted, there is also the larger ethical issue.

[Chataignier]

We are getting off topic, but I do agree that there are limits to consent. In the standard form of consent I use in France it states :

Le photographe s'interdit expressément de procéder une exploitation des photographies susceptible de porter atteinte à la vie privée ou à la réputation, ni d'utiliser les photographies, objets de la présente, dans tout support à caractère pornographique, raciste, xénophobe ou toute autre exploitation préjudiciable.

For those who don't read French, it says (amongst other things) that the photographer agrees not to use the image in any way that is prejudicial to the subject.

[Round Tuit]

"I own the copyright on original image and I can do whatever I want with it."

Sorry didn't know that.

You could not know that because it is not true; at least not in Canada.
A few years ago I sought a legal opinion on the current copyright law in this country. In a nut shell:
1. You do not need permission to take a photo of anything or anybody that can be reasonably seen from a public place.
2. As the photographer, you own the copyright for the photo. You therefore have exclusive rights to use or publish it as you see fit. You can therefore grant licences for specific or general for free of for considerations.
3. You cannot use the photo of a person for commercial use without the explicit consent of the person.
4. You cannot use the photo in a manner that is prejudicial to the "model(s)".
5. Anybody can use your copyrighted photo without your permission for "educational purpose" in a non-commercial setting.
Those are the broad lines of the opinion that I got. There was also several pages of "interpretations and caveats" so I would strongly recommend seeking legal advice if you are in doubt.

[Astro]

This is an interesting thread, but I see the whole issue probably a bit naively in the context of the following proposition...

I am a fairly incompetent artist... at best I can produce a technically reasonable sketch.
I have a very good friend who is a very accomplished artist and in truth makes my attempts pitiable.

So if I start a sketch and then ask her to finish it for me ????

I may 'own' the final result but I cannot claim any 'creative' rights and being honest about it I would be embarrassed to even consider claiming it as 'my' work.

I have to agree with Bill, and this for me is a step too far when it comes to me expressing my 'creative' idea's. My immediate facile opinion is that such images are best considered Frankenstein creations.

(As a tool in the marketing advertising etc toolset I have no issue with use of this application, mostly because I am a cynic and don't look in general, for artistic 'authenticity' or truth, with advertising/marketing these days!)