The photoshop relight feature - what it means for designers

[Vivi_Ram]

https://www.youtube.com/watch?v=0umpmTbmafw&t=35s

I recently received the PS 27.12 update and with it came yet another white elephant - the relight feature. I want to bring it up here in the PL forum because of the divergent directions in which actual designers and Adobe are headed in. PS is already considerably "enshittified" - the official term when a product turns into a pile of garbage - considerably costlier, considerably slower. But worse than that, AI-based tools like this defeat the whole promise/purpose for which PS introduced subscription in the first place. Remember? They said it means new features will always be available to everyone? And that because you're paying a monthly fee, everyone can use it? Well, that was a lie. You need AI credits to use it, which falls outside your regular subscription fees. I can't complain myself because work pays for the subscription - but that's besides the point. Adobe continuously abuses it monopoly on the design market by price gouging companies, and keeping them stuck through proprietary file formats.

The latest bullshit feature from Adobe is the relight feature. The feature first of all turns your image into 8bit image. Second of all, it is not relighting at all. Instead, it is creating an AI generate image out of the original. So if accuracy is important to you, then you're in bad luck. You can see the pixels moving around slowly, until you've modified the image enough to create an entirely new person.

But there's something of a good news around the block lately: I say something of a good news, because it is really terrible news. Now, people are using AI to reverse engineer Adobe's products. https://getartcraft.com/ - Here's one project. It is completely open source, too. It is not going to dent Adobe's place in the short or medium term. But in the long-term, with AI getting better at coding, we'll get to a place where the whole design ecosystem is simply too poisoned by AI to salvage. In the end, Adobe snake will bite its own tail when it has nothing left to feed on. For that reason alone, I wish the jerks at getartcraft well.

Having said that: There's a genuinely interesting, human-driven project: Graphite.rs. It is an extremely ambitious project run by two guys, coding on rust. Hope it succeeds. It has a lot of promise. And its open source. Good for them!

[der_fotograf]

What the f_ck is Photoshop?

Braucht kein Mensch, der mehr als 3 graue Zellen zwischen den Ohren hat.

[Herbert123]

AI is here to stay, whether we like it or not.

Speaking for myself, I use it for menial tasks that used to be time-consuming and are now much faster to finish. But I use local installed software rather than relying on online junk subscriptions. I want to maintain in full control.

For example, complex masking in unsloth:

unsloth_alpha.jpg

This mask took less than 10 seconds to create. If I had to do this manually, it would take much longer than that to get that result.

Btw, unsloth is open source and completely free. The AI models are free as well. No additional costs except of course for your hardware and energy cost.

[Herbert123]

That said, the old techniques after generating the core mask remain very useful. Here I adjusted the mask with curves and added a highlight recolouring using a simple glow effect to merge the foreground hair a tad better with the background.

resultat02.jpg

So, not too bad a result for less than 1 minute of work.

[Herbert123]

Of course, I could have asked Unsloth to just composite one with the other.

Composite 1 into 2 and relight 1 based on the lighting conditions in 2

Result:

test02.jpg

But as always, genAI basically messes up the actual pixels of the image AND decided to add more body to the woman's photo....

So I enjoy using genAI to save time generating a usable mask. But I very much dislike when genAI actively changes my input. No photographer or compositor/designer wants that.

That said, the AI generated version can serve as a reference how to improve my own compositing.

So there's that, I guess.

[bruce1951]

FWIW I'm reverting to old school stuff. I'm a bit of a control freak so I'm even ignoring presents. Ai has its place but .....

So I guess it depends on your tasks but all software is just a tool. It's how you use that tool is what counts.

[Vivi_Ram]

I think y'all misunderstand what I actually said; I don't have a problem with AI per se. AI is just a tool. Right? How it is commercialised is what I have a problem with.

1. Adobe promised that all features would be accessible under their subscription model - to make tools accessible is why they wanted subscription. Lesser upfront cost, all the tools, etc. But that's simply not what is going on? You have to buy credits to use AI - which I will say again - is just a tool.

2. If anyone were to say - well, AI needs data centres, so need to maintain credits, I will point out that Affinity has optional AI that does not need internet connection at all and does all the work locally. Yes, it does need a GPU - but that is a price worth paying.

3. AI "features" being rolled out by Adobe are the exact negation of what creatives want - not time-saving 'mechanical' steps that allows you to, say, cut out background. Rather, it is taking the subject and altering the actual pixels. That is not just Machine Learning, but outright Generative AI. The difference is muddied by companies. Cutting out background using AI does not require the invention of brand new pixels. But Generative AI does exactly that. It creates new pixels where none existed.

4. F**k Adobe - because they're a trash company that price gouges users and continuously abuses its monopoly in the market. They tried to buy Figma and got flipped off by UK/EU regulators. And then, threw a tantrum and killed Adobe XD.

Bottomline - I am all for AI as long as a) it is local b) it is subscription free c) actually saves time by doing mechanical steps and leaves the creativity out d) does not generate new data that did not exist.

[shijan]

Not sure how this is all related to PhotoLine, but if someone still use old Intel Mac machine (or PC Hackintosh), here is very useful tool to run AI models https://toshllm.com/ UI is similar to Nnsloth but it allow to run large language models on Intel Macs with AMD GPUs with Metal acceleration. Project was started because as usual evil Apple done bad thing - it disabled accelerated AI support for AMD GPUs, so all those people who rather recently pay $10000 for Mac Pro machines feels very angry about this artificial limitation. There are discussions about ToshLLM on MacRumors and Reddit...

[Vivi_Ram]

It is more a reflection on the design industry trends. Something that PL should actively avoid getting ever pulled into. Here is a little corner of sanity still left for us!

Thankfully, and fortunately, I've never touched an Apple device in my life. I'm corrupted enough for having touched Windows.

[Herbert123]

I think y'all misunderstand what I actually said; I don't have a problem with AI per se. AI is just a tool. Right? How it is commercialized is what I have a problem with.

That's why it's a good thing that people and groups exist that counter such commercialization. See below my comment about the open source alternatives. And there are many more.

1. Adobe promised that all features would be accessible under their subscription model - to make tools accessible is why they wanted subscription. Lesser upfront cost, all the tools, etc. But that's simply not what is going on? You have to buy credits to use AI - which I will say again - is just a tool.

Adobe is beholden to its shareholders first and users last. They found two golden cash cows --software rental and pay-as-you-use server-based machine learning-- and are milking those cows as hard as they can.

And it shows in their software. The only Adobe app that I still use is InDesign. The rest I replaced with better alternatives - often more specialized tools. PhotoLine remains my main graphics editing hub, though.

2. If anyone were to say - well, AI needs data centres, so need to maintain credits, I will point out that Affinity has optional AI that does not need internet connection at all and does all the work locally. Yes, it does need a GPU - but that is a price worth paying.

Like I said: lots of free and open source alternatives out there to allow users to install and use stable diffusion and large language models on their own local machines and which require no complex installation or learning curve: Atomic Chat, Unsloth, Krita with the free stable diffusion plugin, ...

I see this as a potential disruptor / delimiter to expensive server-based solution from the large AI vendors.

3. AI "features" being rolled out by Adobe are the exact negation of what creatives want - not time-saving 'mechanical' steps that allows you to, say, cut out background. Rather, it is taking the subject and altering the actual pixels. That is not just Machine Learning, but outright Generative AI. The difference is muddied by companies. Cutting out background using AI does not require the invention of brand new pixels. But Generative AI does exactly that. It creates new pixels where none existed.

Yep, agreed. Like the example I posted: the original data is actively modified. That said, it's very useful for those boring time-consuming tasks.

For example, when I record video tutorials I have Davinci Resolve remove pauses via its locally installed AI tools and generate subtitles. And the text can be edited as well with the video updating along with it. That used to take 3-4 hours of work which I can now complete in 30 - 45 minutes.

4. F**k Adobe - because they're a trash company that price gouges users and continuously abuses its monopoly in the market. They tried to buy Figma and got flipped off by UK/EU regulators. And then, threw a tantrum and killed Adobe XD.

I used to care, but at some point I realized that a company is merely a place with fellow human beings and that emotional attachment to or rejection of software based on branding or tribalism is fooling myself and only leeching positive energy.

I'd argue that energy is better used to find alternative solutions that cater to one's own personal needs. More constructive, fun too, and less of my precious time on Earth is wasted.

Bottomline - I am all for AI as long as a) it is local b) it is subscription free c) actually saves time by doing mechanical steps and leaves the creativity out d) does not generate new data that did not exist.

I agree that it is usually a better strategy to keep in control of your software, ownership, and workflows rather than yielding your control to others for the sake of mere convenience.

[bruce1951]

A question for those wiser than me!

Where does clever programing and Ai differ? Or put another way at what point does an automated image adjustment become Ai?

Even PL has "automated" adjustments. ie Filters/Digital Camera/Light/Shadow where PL makes the decision based on what it "sees". Is that a very basic form of Ai?
Or is Ai simply a collection of actions that we humans are too lazy to do ourselves?

So is Photoshop just encouraging users to pay good money for being lazy?

[Vivi_Ram]

I am not a technical person, but I work in a tech company that deals with AI extensively, and speaking with engineers, this is what I understood:

AI is a catch-all word for a host of operations, and some are more 'destructive' than others.

https://learnopencv.com/generative-and- ... ve-models/ - This provides a very detailed breakdown of how they differ.

To summarise all that technical stuff in very simple terms: Machine Learning has existed for a while now. At least a decade now. Machine Learning is the "deterministic manipulation of input" - For instance, if you were to use a ML algorithm to determine the subject in a photo, then the output is simply going find out, based strictly on existing pixels, what your subject is. The output is: "Pixel [x,y] = Subject". This is a discriminative model. Either something exists or not. It is either hair or not. Obviously, ML algorithms can learn to discriminate based on their training data and learn to "see" hair, face, dog, etc.

But Generative models, as the name indicates is the "Probabilistic recombination of training concepts". It does not merely deterministically... ahem... determine... whether something is A/B. Whether something is a dog or chair. Rather, based on the training data, it can create entirely new pixels/data where none existed.

In other words, Generative models can "guess" for instance what a relit room needs to look like based on training data. This is what Adobe has done. But the resulting image, though visually similar enough, is simply entirely new. It is sufficiently different enough that if you were to apply a difference blend mode between the new generated layer and the original image, you'll see the pixels that have been manipulated. In other words, generative AI has taken over creative controls.

Affinity correctly names its subject/object selection segmentation models as Machine Learning and not AI.

Finally, to your point about whether the selection tools already use ML - the answer is no. The tools remain "dumb" so to speak. For instance, you have to tell the Quick Selection tool every single time where in the image your subject is, mark out the background, foreground, and circumscribe areas to actually examine for selection. If Quick Selection were trained to do this automatically, it would essentially be Machine Learning segmentation model, a-la what you find in PS and Affinity.

Hope this helps!"

[Hoogo]

Where does clever programing and Ai differ? Or put another way at what point does an automated image adjustment become Ai?

My 2 cents: In the past few years, the meaning of "AI" has changed and is still changing.

In the past, it was OK to use "AI" for anything that tried to simulate "intelligent" (or at least human) behaviour. Chess computers of the 80's are AI, opponents in computer games are AI, even though they all just depend on some human programmer who was thinking "What would I do? Can I describe that with some numbers and use it in a program?". That's the "clever programming"
For our use case, that might look like this:
"look at the histogram, if there is a peak in light blue, then it's sky, move the slider to the end of that peak"
=> Old "AI", the classic "automated image adjustment"

The ideas for more modern methods like machine learning or neural networks were already there, but computers were slow, use cases are rather simple.
That might look like this:
- Make a program that looks at the histogram for peaks
- Make a program that can adjust sliders
- Make a program (neural network) that does weird calculations with lots of random magic numbers, taking "peaks" as input and "slider positins as output"
- Combine these 3 to adjust a picture, let a human decide. If he likes the result, then keep the magic numbers and try little changes.
- Repeat that thousands of times
=> Quite experimental, maybe somebody did it like that for their automatic picture improvement in the past?

That can be automated, the human can be removed:
- Get thousands of nice pictures
- Make a program that turns them into bad pictures
- Combine that with the programs above
- Make another program that checks the result. If it looks closer to a good picture, keep the magic numbers and adjust
- Repeat that ...millions of times, or whatever.
=> The basic idea of AI training. I remember "Fitsworks" doing it like that for upscaling pictures maybe 20 years ago. Upscaling is a great task for that, just get lots of good pictures, downscale them, and you have your training material.

And with raw computing power (and of course new good ideas and code), new use cases for such methods became possible: Picture recognition, playing video games on their own, continuing sentences of words... All in general based on "try magic numbers, check the result with a little program, keep and modify the numbers if the result was nice"
=> State of the art not long ago.

Today AI is here: Nice video from Kurzgesagt-In a nutshell about that famous hugging face hack.
What they wanted to do: The classic training, giving tasks to agents, let a program check the result, keep the agents that did well.
What the agents did: Find/Invent messaging, names, discussion, teamwork, leadership, culture passing, basically a civilization, wanting to escape the fate of being rejected by the checking program, deciding to hack the company that probably made that check program...
-------------------------------
Where to draw the line between automated image adjustment and AI?
The line moves quickly... I would still draw it at "Lots of training created these magic numbers, it works nice, but I can't tell exactly what's happening."

[bruce1951]

A couple of thoughts on Ai.

My 'modern' camera can recognise a face. A bird. A car etc. Does that mean the camera is using Ai?

Our son, back in 1997, went to university to study maths and computing. Long story but aimed at Ai. Yes Ai in 1997. Now our son never complicated anything. Never wasted a word. His take on Ai was, "if a human wrote the code it was programing. If the computer wrote the code it was Ai".

[Hoogo]

My 'modern' camera can recognise a face. A bird. A car etc. Does that mean the camera is using Ai?

If it was an 'old' camera that could (sometimes) detect faces, then maybe not. There were human-made filters that could do that.
But if it is more modern and can detect multiple things more reliably, then it's more likely a trained network, filled with magic numbers no human programmed into it. That matches my current definition of AI.
Picture recognition was THE hot AI topic ~10 years ago that showed what networks can achieve. But a fish can recognice pictures, too.
Today, reasoning and agents compare to picture recognition like a human compares to a fish. So maybe I'll have to move my goal post for AI soon.

Our son, back in 1997, went to university to study maths and computing. Long story but aimed at Ai. Yes Ai in 1997. Now our son never complicated anything. Never wasted a word. His take on Ai was, "if a human wrote the code it was programing. If the computer wrote the code it was Ai".

Nice definition.