AI Assisted Edits Do my photos even need me?

[clee01l]

Operating System

1. macOS 27 Golden Gate

Here are screen shots from an image in CLassic and the same RAW file opened directly into Photoshop and using the new AI Assisted edit process. In Photoshop I told the app I wanted to enhance the subject reduce shadows and diminish the background.
Edited in Classic

AI assisted Edit in Photoshop.

 

[johnbeardy]

Maybe that stubby branch means you forgot to tell it not to add anything? Negative constraints seem important, at least for now.

 

[clee01l]

Maybe that stubby branch means you forgot to tell it not to add anything? Negative constraints seem important, at least for now.

This was a learning experience. I could have added a remove instruction to the earlier instruction as well as crop dimensions and other things. Each iteration uses 30 credits and I did not want to waste another 30 credits to remove the distractions. The Lightroom crop window is 1990X1120. The Firefly image covering approximately the same area is 5056X2844 due to upsizing.

 

[Swingman]

@clee01l , May be you don’t even need a photo to produce the results.

 

[Replytoken]

My two cent takeaway is that AI tried to create an image based on what you gave it as a sample. Look at the wings in both images. They look quite different.

I am not saying that the AI image is not nice, but it barely resembles the photo that you took. Now if your style was to radically edit images that you take, then maybe this is a tool that you might want to consider. But I have no interest in it. I take photos because I like the challenge of taking them, and the challenge of editing them. This seems to negate the latter by the traditional definition of editing.

--Ken

 

[Swingman]

I have been watching a few mins of every session of `Photoshop Virtual Summit` . I couldn’t bring myself to see even one full episode as they all deviate into full AI adventures. They should have called it AI Adventures Summit 2026.
Basically you don’t need a good camera or an operator behind it. We will do everything by choosing a Firefly or a partner model and tell us in few words what you need.
Probably in 2027 summit, your thoughts would be immediately transformed into new artwork.

 

[clee01l]

My two cent takeaway is that AI tried to create an image based on what you gave it as a sample. Look at the wings in both images. They look quite different.

I am not saying that the AI image is not nice, but it barely resembles the photo that you took. Now if your style was to radically edit images that you take, then maybe this is a tool that you might want to consider. But I have no interest in it. I take photos because I like the challenge of taking them, and the challenge of editing them. This seems to negate the latter by the traditional definition of editing.

--Ken

I disagree. the Images while rendered differently are the same. Look at the stem of the plant the Dragonfly has lit upon. The biggests difference is the background whcih is rendered exactly as I requested. I have studied enough dragonflies to know the wing venation is correct. AI would have been creative with the wing veins. I've used Topaz "Super Focus" on other dragonflies. That Topaz AI process invents it own wing venation

 

[Replytoken]

I disagree. the Images while rendered differently are the same. Look at the stem of the plant the Dragonfly has lit upon. The biggests difference is the background whcih is rendered exactly as I requested. I have studied enough dragonflies to know the wing venation is correct. AI would have been creative with the wing veins. I've used Topaz "Super Focus" on other dragonflies. That Topaz AI process invents it own wing venation

I guess I am focusing on the dark spots along the wing edges, and perhaps the background editing may have obscured them. They appear to be more diminished in the second image, and that caught my attention. This could have happened with human editing as well, but I suspect I would feel the same about that too. I'll give you this, they may not look quite as different as I had initially stated, but I do not care for the changes, as I tend to like a light touch when editing. I guess that beauty is in the eye of the beholder.

--Ken

 

[clee01l]

One thing I would like to see with Assisted editing is the edit history the Photoshop took to get the result. Maybe if this Process shows up in Classic, the edit seps will be preserved and possibly reversible.

 

[johnbeardy]

You should find the prompt in the new layer’s Properties panel, and from there you can paste it into a field, either via Photoshop’s File Info or in Lightroom. I have been using the Instructions field to record prompts.

Also remember the History Log which is off by default. I'd be surprised if it records much of this AI stuff, but it might be possible to get log info into Lightroom by a plugin that reads a text file or a plugin that uses Exiftool to access the metadata, or just cut and paste.

 

[RoyReed]

The problem is that the AI has added all sorts of detail that isn't in the original photo.

I asked Copilot to identify the dragonfly before and after. It said before looks like a Blue Dasher (Pachydiplax longipennis), after as a Western Pondhawk (Erythemis collocata) or Eastern Pondhawk (Erythemis simplicicollis). I got similar answers when I asked Gemini (although it admitted it could be wrong). Fine if all you want is a pretty picture, but not if you want a photograph of a particular species.

 

[clee01l]

The problem is that the AI has added all sorts of detail that isn't in the original photo.

View attachment 30453

I asked Copilot to identify the dragonfly before and after. It said before looks like a Blue Dasher (Pachydiplax longipennis), after as a Western Pondhawk (Erythemis collocata) or Eastern Pondhawk (Erythemis simplicicollis). I got similar answers when I asked Gemini (although it admitted it could be wrong). Fine if all you want is a pretty picture, but not if you want a photograph of a particular species.

Yes now I do see some added features.. In particular the antennae which are not visible in the original.

I had previously identified this specimen as Great Blue Skimmer (Libellula vibrans). I submitted the images to Claude and also used my phone "Picture Insect" app and both agreed with my original identification. which I'm sure used other views and perhaps live specimens. So I am confident that my 2009 identification is the correct one.

 

[RoyReed]

Yes now I do see some added features.. In particular the antennae which are not visible in the original.

I had previously identified this specimen as Great Blue Skimmer (Libellula vibrans). I submitted the images to Claude and also used my phone "Picture Insect" app and both agreed with my original identification. which I'm sure used other views and perhaps live specimens. So I am confident that my 2009 identification is the correct one.

Both of the AIs I used mentioned Great Blue Skimmer, but thought their main suggestions much more likely. Maybe more precise location data than just 'USA' would have helped. I know that identifying UK dragonflies and damselflies can be equally problematic.

 

[johnbeardy]

Did your prompt include instructions not to change any of the details of the insect?
Generating lengthy negative constraints is why I use Copilot to write full prompts.

 

[clee01l]

Did your prompt include instructions not to change any of the details of the insect?
Generating lengthy negative constraints is why I use Copilot to write full prompts.

No they did not. And I am beginning to see why that might be necessary.


Sent from my iPad using Tapatalk

 

[johnbeardy]

I feel it's less necessary than 18 months ago when I first started experimenting, and maybe in time it won't be needed in certain apps?

One advantage of keeping the prompt in Instructions is that I can always try them again and see if things have improved.

 

[Jimmsp]

All AI models are not created equal, and because diffusion and vision-generation models are probabilistic, running the exact same prompt twice won’t produce identical pixels.
What's happening here highlights the difference between generative re-creation (which tends to invent or hallucinate fine structures like antennae and wing venation) and true local editing.
OpenAI just released ChatGPT Images 2.5 (and their GPT-Image-2.5 models), which specifically focus on preserving the subject from reference photos and isolating edits to specific areas like the background without rebuilding the subject. John is spot-on: detailed prompt instructions and explicit negative constraints remain essential.

 

[johnbeardy]

running the exact same prompt twice won’t produce identical pixels.

For pixels, it's worth adding "looks" or "styles".

I think a lot of photographers produce groups of pictures, and they want an entire set or shoot to have common styling. You want to apply the same prompt like "make it look like 1980s cross processing..." to a series of images and get results that look like a family, but too often they're such different interpretations of the same prompt that you wonder if an AI clinic got its samples all mixed up. Lengthier prompts and restrictions take out the guesswork.

 

[UtahLlama]

As far as I'm concerned "AI Assisted Edit" is disingenuously named. It's not Editing in the traditional sense of moving sliders, masking or cloning, it's not even Assisting you with suggestions of how you might edit. It's just drawing you a new image. Your original photo is acting as prompt for content and style, a picture is worth a thousand words, as they say.

 

[clee01l]

As far as I'm concerned "AI Assisted Edit" is disingenuously named. It's not Editing in the traditional sense of moving sliders, masking or cloning, it's not even Assisting you with suggestions of how you might edit. It's just drawing you a new image. Your original photo is acting as prompt for content and style, a picture is worth a thousand words, as they say.

I don't disagree. However, it is AI that is doing the editing and the result is still an "edited" image. After AI has produced an edited image, you are still able to apply manual adjustments.