Part 3 of 3: AI and photography
I’ve used AI coding assistants to build photography tools by describing the problems, testing the results and feeding back what needed improving.
Quick answer: I built tools for colour analysis, predicted visual attention, image descriptions and face-similarity comparisons using Codex and Claude Code in separate projects. I didn’t write the code myself. My contribution was knowing the editing problem, explaining the result I wanted and testing whether the tool was useful. These are tools I’ve built, rather than products I’m announcing for sale.
Why did I start building photography tools?
If you use Photoshop, you’ve probably seen editing panels advertised.
They’re little extra interfaces full of buttons and controls, built to make particular jobs easier. I’ve bought panels myself, including TK Luminosity Masking and Lumenzia. They gave me ways to work that would otherwise have involved more manual effort.
I always assumed making something like that belonged to people who could code. A photographer would have an idea, get a developer involved and work out whether enough people would buy it to make the exercise worthwhile.
Then I found myself doing something I hadn’t expected: building tools for my own photography work, without writing a line of code.
It started with the things I found awkward. Colour theory was one of them. I’d watched videos, and I understood parts of it, but applying it still felt like a bit of a dark science.
Rather than watch another explanation and hope it finally stuck, I could describe the thing I wanted help with and work on a tool that made it easier to see.
What did I actually have to contribute?
The problem still had to come from me.
I had to explain what I was looking at, what I found difficult and what would make the task more useful. Then I had to try the result and say what worked, what felt awkward and what I wanted to improve.
That is a different skill from writing code, but it’s still work. You need enough understanding of the task to judge the result.
It’s related to what I wrote about in giving AI useful context about your business. A vague request leaves an awful lot for the AI to guess. Explaining the real problem gives it something more useful to work on.
For these projects, the context was photographic: an image, an editing frustration, and an idea of the controls I would want to use.
What do the colour tools do?
I’ve built two panels concerned with colour theory.
They can analyse the dominant colours in a photograph, help me select colours I want to adjust and suggest colour schemes, such as complementary colours.
That makes an abstract subject more practical for me. I’m looking at my own photograph and seeing the colours I’m dealing with, rather than trying to remember which bit of a tutorial applies to it.
The suggestion doesn’t have to become the finished look. It’s something I can consider and work with. The photograph still has to look right to me.
Can a tool show what attracts attention in a photograph?
I’ve also built a panel that uses an AI model to create a heat map predicting visual attention.
The question I want it to help with is quite simple. Am I drawing attention to the person, or is the shiny thing in the background winning?
The panel also creates masks and curves layers. Those give me a way to brighten or darken areas after looking at the prediction, rather than just staring at a colourful diagram and wondering what to do next.
The word “predicting” matters here. I haven’t measured where a roomful of real people looked. The model is offering an estimate, which I can use alongside my own judgement.
That’s still useful to me. I just wouldn’t describe it as proof of what every viewer will notice.
Why build an image-description tool?
Another tool uses a local AI model to analyse a photograph and write text descriptions at different levels of detail.
One possible use is helping prepare descriptions for photographs online, where someone using a screen reader needs information about what the image shows.
The tool gives me text to work with. That doesn’t mean every description is accurate or suitable to paste straight onto a website. The useful result is a starting point I can review, rather than a reason to stop checking what the photograph actually contains.
What is the face-comparison tool for?
I’ve built a tool that compares a generated likeness with a photograph of the actual person and returns a similarity score.
The idea is to get an indication of how close the likeness is. A generated face might look convincing as a face while looking very little like the person it is supposed to represent.
Again, a model score is an indication. It isn’t proof of identity, and I haven’t established a magic threshold that guarantees a good likeness.
I find this sort of project interesting because it answers a specific question I’ve had while working. It gives me another way to look at the problem.
How do I go from an idea to something useful?
My process is fairly practical:
- Describe the frustration. What takes too long, or is difficult to understand?
- Explain the result I want. What would I like to see or control?
- Work on a first version with the AI. Give it enough context to build something relevant.
- Try it. A tool needs to work for the photographic task, not just open successfully.
- Give feedback. Explain the awkward bits and suggest improvements.
- Try it again. Judge the result by whether it helps with the original problem.
The testing and feedback are a big part of it. An idea can sound good when you describe it, then need a different arrangement of controls once you’re using it.
Does this make photography knowledge less important?
In these projects, it’s what made the tools possible.
I knew enough about the editing problem to describe it. I knew which parts I found frustrating. And I could look at the result and decide whether it helped.
AI gave me a way to turn that knowledge into software, without learning to write all the code first.
That’s the part I find most useful. Ideas that used to stay in my head can become something I can open, try and improve. Some help with colour, some help with analysis, and some take a repetitive job off my hands.
They’re one more part of how I work on photographs for clients. The decisions about what the photographs need to do still come first.
Cheers,
Ade
Frequently asked questions
Did you write the code yourself?
No. I used Codex and Claude Code in separate projects. I described the problems and desired results, then tested the tools and gave feedback on usability and improvements.
Are these tools available to buy?
This is an account of tools I’ve built for my work, rather than a product launch or an offer to buy them.
Does an attention heat map show where people actually looked?
No. My panel uses a model to predict attention. I use that estimate to help assess an image, alongside my own judgement. It isn’t a record from an eye-tracking study.
Is a face-similarity score proof that a generated portrait looks right?
No. It offers an indication from a model. It doesn’t replace looking at the portrait, and it isn’t identity verification.