I am discovering some actual benefits of A.I. for legitimate image post processing

[peterm1]

I have posted this here rather than in the post processing forum as that forum seems to focus on film, Lightroom etc. - i.e. more traditional tools. So I am treating this as a general discussion on AI mediated PP not on techniques etc.

So, over the past week or so I have been experimenting with A.I enhanced image processing using (for the most part) Google Gemini a free A.I. based tool (or perhaps mostly free - I have not explored this aspect yet and suspect more advanced tools are available at a price). I learned of this through a Facebook group dedicated to posting vintage images (mostly 19th and early 20th century) of my home city, Adelaide, South Australia. In the post he advised members how the software can be used to enhance and colorize etc. very old photos. I tried this with some very old family photos and was impressed, then the thought struck me that I have some older photos made by myself, mostly back in the early digital era where (a) I had no clue what I was doing and (b) the sensors and firmware were much more primitive than it is today. In particular some images had very low resolution - not quite literally 'postage stamp" size but not far off either, very frequently excessively grainy, quite often blurry (my fault mainly), having poor exposure and generally beyond redemption using traditional post processing software and techniques because too much of the subject was missing from the captured image due to the above flaws.

I decided to try to salvage some of these disasters and indeed in many cases (though not all) it proved to be possible to do so with results which in some instances were pretty damn good. Admittedly, I noticed that in some cases where the starting images were too poor, I could see that the A.I produced results did not look perfect because it was being forced to make elements of the image up out of whole cloth (as they say). I noticed this occasionally, especially in portraits where the end result looked ever so slightly "off" in terms of proportions of a face etc. (Bear in mind that A.I sometimes has to make pixels up from thin air and best guesses to fill in blanks.) And in some cases I felt the result looked a little artificial or even clinical and hence a bit "plasticky". Nevertheless surprisingly often I was singularly impressed with results. For now, I will just post two of myself so you can see for yourself how this works (or not). The starting image in this pair was made quite recently actually but the image was of me at the computer and was taken by a P.C.'s built in camera in a darkened room. As might be expected the result was pretty poor. I directed the AI to enhance detail, remove artifacts, improve tonal values and saturation and enlarge the image. I was particularly impressed with how the A.I. responded to my request for it to remove the flare in my eye-glasses.

I still have some reservations about how far an AI processed image can be taken before it can be regarded as a wholly unethical outcome. But I can certainly see promise in these tools' ability to fix problems that were otherwise beyond my ability to fix and maybe beyond the ability of anyone to fix by any other means.

One afterthought. I tried much the same on a scene from a photo of Budapest made around 2000. It was shot on Ilford XP-2 (black and white) and later scanned on a flat bed scanner, probably at fairly low res. One instruction I gave the AI was to colorize the image in "colors suited to late Spring or early Summer in Budapest, Central Europe" While the colors resulting were a little painterly, damn me, the choices made by the AI tool were pretty acceptable. A lucky guess? Time will tell I suppose if it can be replicated with other locales. But it left me wondering, is this thing clever enough to know what Budapest in late Spring looks like?

In the first photo below the final result appears first.

 

[raydm6]

Looks good @peterm1. I think there are positive uses for AI in photography and in post-processing.

As far as ethics, I just completed a free Google AI Essentials Specialization Certificate program (which also uses Google Gemini for class work although you are not limited to that AI tool) and finishing up another certificate: Google Prompting Essentials Specialization.

In both paths, ethics and respect/inclusion of others is mentioned quite often along with the realization that these tools are still relatively new, error-prone (AI Hallucinations), and always require human-in-the-loop interaction and approach during prompt design and evaluation/iteration.

Google training (along with others) provides a very good prompting framework - Task, Context, Reference, Evaluation, Iteration - to get the most out of AI tools - again, stressing the importance of the human as the final arbiter of output. Folks might find that as a useful approach when designing prompts.

There's a big push to leverage AI at our workplace. We are authorized to use Microsoft Copilot AI which uses OpenAI's GPT-5.2 LLM. So far, I have found it useful to analyze large Excel & .csv file data-sets (telemetry data from over >10k devices) and reports (I received three different device failure analysis reports yesterday and Copilot was able to give me a brief tabled summary and cross-unit comparison very quickly before opening and diving deeper into the reports). Quite powerful and compelling.

Still learning how to use it in a more deeper fashion however...

 

[JohnWolf]

.... Google training (along with others) provides a very good prompting framework - Task, Context, Reference, Evaluation, Iteration - to get the most out of AI tools - again, stressing the importance of the human as the final arbiter of output. Folks might find that as a useful approach when designing prompts....

This is a good point. No doubt some of the disappointment with AI query results is because of poor quality prompts. Besides Google's, there are a number of prompting frameworks and they are worth learning about. I was taught the acronym "CO-STAR" - Context, Objective, Style, Tone, Audience, Response. Once you start using them, the results improve significantly.

 

[splitimageview]

Yes results are mostly a matter of prompt quality (but not always..,I sent a photo of an M2 to ChatGPT yesterday and for no reason at all it turned the serial number upside down.)

I’m a member of a ‘growing up in the 70s’ group and the last few weeks have been nothing but submissions of old faded or torn photos, asking for assistance to resurrect something usable.

Other members with decent prompting experience then throw it at AI.

I’m not sure some of the members realize what’s happening, they might think others are using old-school Photoshop skills. 😀

 

[peterm1]

I thought I would throw this photo up as well. I mentioned it in my opening post in which I took a black and white image and sought to colourize it. While the image as shown here looks a little too much like a picture postcard of old, bear in mind that my instructions were very brief - "Apply colors suited to late Spring or early Summer in Budapest, Central Europe" and also to reduce artifacts in the image. Maybe there was a further iteration to fix something or other and I also feel I could get it to apply more AI mediated sharpening and / or detail enhancement but have not done so. Overall I am pretty happy with the result for such a brief experiment and it should be possible to tweak it further in Lightroom to make the colors less painterly if that is what I decide to do (I may not as this was primarily not much more than an experiment in how to use the AI and whether it would make sensible choices as to colors with minimal guidance.)

 

[Freakscene]

Yes results are mostly a matter of prompt quality (but not always..,I sent a photo of an M2 to ChatGPT yesterday and for no reason at all it turned the serial number upside down.)

I’m a member of a ‘growing up in the 70s’ group and the last few weeks have been nothing but submissions of old faded or torn photos, asking for assistance to resurrect something usable.

Other members with decent prompting experience then throw it at AI.

I’m not sure some of the members realize what’s happening, they might think others are using old-school Photoshop skills. 😀

My 12 year old son and I have been assisting a lot of gen xers and older millennials to do this too.

 

[peterm1]

My 12 year old son and I have been assisting a lot of gen xers and older millennials to do this too.

"
My 12 year old son .......................... assisting a lot of gen xers and older millennials to do this too.


Yeh thanks, that really makes me feel good. 😆 😆 😆 😆


Though I see your point!

And perhaps I can take heart from the fact that as a "boomer" I am older than the named groups and hence perhaps ahead of my gen.

 

[Freakscene]

"
My 12 year old son .......................... assisting a lot of gen xers and older millennials to do this too.


Yeh thanks, that really makes me feel good. 😆 😆 😆 😆


Though I see your point!

And perhaps I can take heart from the fact that as a "boomer" I am older than the named groups and hence perhaps ahead of my gen.

My son did some robotics and programming courses in primary school and is already competent at programming in C++. I showed him the statistical analyses that I do for work in R, and he thought about it a while, and said about R “It’s nice and lean, but you still need a program to run it. I could just build the whole thing in a construction language and run it in terminal”. He is a total digital native, and sometimes his knowledge and capacity is breathtaking.

 

[peterm1]

My son did some robotics and programming courses in primary school and is already competent at programming in C++. I showed him the statistical analyses that I do for work in R, and he thought about it a while, and said about R “It’s nice and lean, but you still need a program to run it. I could just build the whole thing in a construction language and run it in terminal”. He is a total digital native, and sometimes his knowledge and capacity is breathtaking.

Yep now you are really making me feel like Sh#t. 😆 😆 😆 😆 😆 😆
Joking of course!

 

[peppard]

My son did some robotics and programming courses in primary school and is already competent at programming in C++. I showed him the statistical analyses that I do for work in R, and he thought about it a while, and said about R “It’s nice and lean, but you still need a program to run it. I could just build the whole thing in a construction language and run it in terminal”. He is a total digital native, and sometimes his knowledge and capacity is breathtaking.

I don't know what to say...reading this makes me feel very old. I graduated in mechanical engineering at the university of Minnesota eons ago, and in my programming computer class we used punch cards, using a reading machine the size of a semi truck. Very happy for you child to be so bright and capable at his tender age.

 

[Jonathan R]

My son did some robotics and programming courses in primary school and is already competent at programming in C++. I showed him the statistical analyses that I do for work in R, and he thought about it a while, and said about R “It’s nice and lean, but you still need a program to run it. I could just build the whole thing in a construction language and run it in terminal”. He is a total digital native, and sometimes his knowledge and capacity is breathtaking.

You use R for stats analysis. Are you a biologist?

 

[Ray Kilby]

I have just taken two images randlomly, fed them into Gemini and asked with the colour street scene, 'Could you improve this?' and with a mates portrait, I asked Gemini "To very subtly make him look younger" These are the results which took me seconds. Crumbs... I'm a little taken aback to be honest. I mean these are just dashed off, no thought. Ummm, thinking. Useful for weddings and corporate portraits and maybe some brochure work where you don't give a damn apart from the fee maybe. I know I won't be using this. It takes the fun out of the post. But damn... I am a little shocked. (I have left the small AI symbol in. the corner for clarity.)

 

[Ray Kilby]

I hope this isn't too off-topic, but I believe that as AI evolves, those in technical roles or fields involving statistics and numbers might actually struggle the most. Conversely, I think creative people will thrive. We could be heading toward a new Renaissance where artists are also scientists; much like Leonardo da Vinci. Maybe it’s just wishful thinking, but I wonder if AI will ultimately become a tool for creatives to build something entirely new, even if we can't define it yet. Perhaps this shift will only happen after we move past the initial 'abuse' of these tools. With any luck, the next decade will belong to the artists. I was born optimistic. Maybe foolish. Is this wishful thinking?

 

[Freakscene]

You use R for stats analysis. Are you a biologist?

I was a medical scientist, more recently a marine biologist by training. I use latent class models a lot for comparison of diagnostic methods and use R for a lit of statistical analysis.

 

[mrtoml]

I was a medical scientist, more recently a marine biologist by training. I use latent class models a lot for comparison of diagnostic methods and use R for a lit of statistical analysis.

I use R too (and Stata). What package do you use for the latent class analysis? I am a sociologist by the way (semi-retired).

 

[Jonathan R]

I was a medical scientist, more recently a marine biologist by training. I use latent class models a lot for comparison of diagnostic methods and use R for a lit of statistical analysis.

I use R too (and Stata). What package do you use for the latent class analysis? I am a sociologist by the way (semi-retired).

I feel I am among friends! I was a wildlife biologist, retired 6 years ago. R became the default for statistical analysis latterly, after many years of Genstat and before that, Systat (and before that, calculating test statistics on a pocket calculator!)

Now back to AI...

I have just taken two images randlomly, fed them into Gemini and asked with the colour street scene, 'Could you improve this?' and with a mates portrait, I asked Gemini "To very subtly make him look younger" These are the results which took me seconds. Crumbs... I'm a little taken aback to be honest. I mean these are just dashed off, no thought. Ummm, thinking. Useful for weddings and corporate portraits and maybe some brochure work where you don't give a damn apart from the fee maybe. I know I won't be using this. It takes the fun out of the post. But damn... I am a little shocked.

Of the street scene, I see little to choose between them. For the portrait, the AI's decision to lower overall contrast, presumably to reduce cragginess, both removes character in the sitter and spoils the overall tonality of the photo. Straightening the smile has removed some complexity of expression that was actually interesting. Lifting the dewlap...well, let's face it, this is a guy whose face is evidence of a life lived, not a youth whose looks are all about potential. Henri Cartier-Bresson said that by the time people are 40 they have the face they deserve. That seems a bit harsh in my case (like, what did I do to deserve this?), but you can see what he was driving at: he found older sitters more interesting.

 

[Ray Kilby]

I feel I am among friends! I was a wildlife biologist, retired 6 years ago. R became the default for statistical analysis latterly, after many years of Genstat and before that, Systat (and before that, calculating test statistics on a pocket calculator!)

Now back to AI...

Of the street scene, I see little to choose between them. For the portrait, the AI's decision to lower overall contrast, presumably to reduce cragginess, both removes character in the sitter and spoils the overall tonality of the photo. Straightening the smile has removed some complexity of expression that was actually interesting. Lifting the dewlap...well, let's face it, this is a guy whose face is evidence of a life lived, not a youth whose looks are all about potential. Henri Cartier-Bresson said that by the time people are 40 they have the face they deserve. That seems a bit harsh in my case (like, what did I do to deserve this?), but you can see what he was driving at: he found older sitters more interesting.

Oh, I completely agree with your observations. I was just trying it out for the first time to see what it could do. As I said, it’s great for weddings and corporate portraits.That was me being a bit 'grand'. I have no right to be, as I did a few of those jobs in the past and I was grateful for the work as they paid well. But the post was soooo boring on all that damned portraiture. The street scene, when enlarged, is very pixelated, those unpleasant shapes are evident and he portrait is better without the Ai enhancement. But I was literally shocked when I tried it out; it’s only going to get better, or worse, from our perspective. I also tried an out of focus shot that I asked it to sharpen. And bugger me, it did it. (Now that is tempting if very dishonest.)

 

[peterm1]

Looks good @peterm1. I think there are positive uses for AI in photography and in post-processing.

As far as ethics, I just completed a free Google AI Essentials Specialization Certificate program (which also uses Google Gemini for class work although you are not limited to that AI tool) and finishing up another certificate: Google Prompting Essentials Specialization.

In both paths, ethics and respect/inclusion of others is mentioned quite often along with the realization that these tools are still relatively new, error-prone (AI Hallucinations), and always require human-in-the-loop interaction and approach during prompt design and evaluation/iteration.

Google training (along with others) provides a very good prompting framework - Task, Context, Reference, Evaluation, Iteration - to get the most out of AI tools - again, stressing the importance of the human as the final arbiter of output. Folks might find that as a useful approach when designing prompts.

There's a big push to leverage AI at our workplace. We are authorized to use Microsoft Copilot AI which uses OpenAI's GPT-5.2 LLM. So far, I have found it useful to analyze large Excel & .csv file data-sets (telemetry data from over >10k devices) and reports (I received three different device failure analysis reports yesterday and Copilot was able to give me a brief tabled summary and cross-unit comparison very quickly before opening and diving deeper into the reports). Quite powerful and compelling.

Still learning how to use it in a more deeper fashion however...

My apologies for not replying earlier - I meant to but somehow overlooked it. Thank you for this, it is very useful information and I may do the course myself. I have certainly observed the importance of good prompting and also how AI can suddenly go off at a tangent and do something quite weird and unbidden. I was using it the other day, for example, to improve the post processing of an image. I was making iterative changes to separate versions of the image figuring this may be more likely to get the outcome I wanted. For the next version I asked for some minor changes to lighting / shadow (about which I was very detailed and explicit) and despite stipulating that there were to be "no other changes, additions or subtractions" in the next version) whereupon the Ai inexplicably (to me, though I suppose there was a reason linked back to poor prompting by me) completely changed the posture of the human subject from standing to kneeling and added two more humans into the image. I had also stipulated the importance of maintaining photo reality with regard to both subject and scene. It's sometimes like dealing with a petulant and impudent staff member who does not know how to take directions. 😆 At other times it surprises me with just how good it is - almost as if it is reading my mind.

 

[Freakscene]

I use R too (and Stata). What package do you use for the latent class analysis? I am a sociologist by the way (semi-retired).

I use R studio. The LCM is mostly done in OpenBUGS using JAGS for the Bayesian analysis, although I now have custom code to do it.

And, most importantly for the future, if you ask any open or contained AI, particularly Claude Code, it can assemble the whole thing, analyze the data and visualize the results. I definitely direct students who only want to use R to obtain their analytical results to follow this approach. I have resources to help them check that the analytical code is correctly compiled and that the outputs are biologically and analytically feasible.

 

[mcfingon]

My apologies for not replying earlier - I meant to but somehow overlooked it. Thank you for this, it is very useful information and I may do the course myself. I have certainly observed the importance of good prompting and also how AI can suddenly go off at a tangent and do something quite weird and unbidden. I was using it the other day, for example, to improve the post processing of an image. I was making iterative changes to separate versions of the image figuring this may be more likely to get the outcome I wanted. For the next version I asked for some minor changes to lighting / shadow (about which I was very detailed and explicit) and despite stipulating that there were to be "no other changes, additions or subtractions" in the next version) whereupon the Ai inexplicably (to me, though I suppose there was a reason linked back to poor prompting by me) completely changed the posture of the human subject from standing to kneeling and added two more humans into the image. I had also stipulated the importance of maintaining photo reality with regard to both subject and scene. It's sometimes like dealing with a petulant and impudent staff member who does not know how to take directions. 😆 At other times it surprises me with just how good it is - almost as if it is reading my mind.

Thanks Peter,
I used your article to try colourizing an old photo and was impressed with the results. It's good to try new things!
John Mc