If you photograph weddings, you probably know the feeling. You’ve photographed the day, ingested the files, and now you need to sit down and look through the whole thing again. That’s what led me to try wedding photo culling with ChatGPT.
Somewhere between the almost-right expression and the slightly-better version of the almost-right expression, there’s gold.
You just have to find it.
Culling is the most tedious part of the job for me. I’d much rather spend that time photographing people or working on the images I’ve selected. But there’s also something else going on when I revisit a wedding: I relive the challenges of shooting it.
I remember the difficult light, the awkward space, or the moment I was trying to make work. Before long, I’m critiquing what I could have done better while also trying to decide which photograph to keep. It’s quite easy to get drawn into that.
Shooting silently on the Sony A1 adds another ingredient. I find that with the lack of audible feedback from the shutter, I shoot more, particularly when I’m working through a moment and waiting for the expression, gesture or connection to come together. That gives me plenty of alternatives—and plenty to sift through afterwards.
So yes, culling can feel rather like panning for gold.
Recently, I’ve been experimenting with a ChatGPT agent to help me find a starting point. We began by working through one wedding together, comparing our selections and discussing the differences. That first cull is the focus of this post, and the photographic examples here all come from that wedding.
This is very much a work in progress. I may decide to go back to doing the whole cull myself. For now, I thought it would be interesting to share the experiment with other photographers, particularly those who haven’t considered using GPT this way or are curious about different approaches.
Why I tried ChatGPT for wedding culling
There’s a reason tools such as Aftershoot and Imagen have found an audience. I’ve tried both, including automated editing, although those experiences were early on and the products may have changed considerably since.
When I tried Aftershoot’s culling, I found I still had too many options. For my taste, the selection wasn’t tight enough to give me the starting point I wanted.
On the editing side, I already have my Lightroom presets set up around how I shoot. My imports usually start fairly close to where I want them, and I fine-tune from there with local adjustments, giving a little extra TLC to the photographs that deserve it. At the time, checking and adjusting automated edits added more work than it saved.
I know photographers who swear by these tools. My experience reflects how they fitted my workflow then, rather than a judgement on their current versions. I haven’t tried Lightroom’s automated culling, so I can’t offer a personal comparison there.
What impressed me about the ChatGPT agent’s first cull was how tight it was. It picked out a surprisingly focused set of photographs before it had even seen any of my finished galleries.
That gave me somewhere useful to begin.
Sometimes it found exactly the frame I wanted. Sometimes it recognised the right moment, but I preferred an expression or gesture a few frames either side. Its selection helped direct my attention, and I could then finesse the choice.
My Photo Mechanic and Lightroom workflow
That fits into a distinction I make in my own workflow: I cull in in Photo Mechanic, then cull out in Lightroom.
In Photo Mechanic, I make a relatively loose selection of photographs worth taking forward. In Lightroom, I edit those choices and remove anything that becomes unnecessary as the gallery takes shape.
Culling with ChatGPT gives me another reference for that first stage. Its tighter selection highlights promising moments, while I can look around those suggestions and decide which alternatives to include.
I still review everything, including the photographs it hasn’t selected. A Red label is a suggestion. I make the final decisions about which moments—and which versions of those moments—belong in the gallery.
How we set up wedding photo culling with ChatGPT
The practical setup gives the ChatGPT agent access to the relevant wedding folders and Photo Mechanic. We’ve given it instructions about how I shoot, examples of my finished work, and feedback on the choices where we differ.
The process looks like this:
- Give it some context. I show it finished galleries and explain what matters to me.
- Let it review the whole shoot. It looks through the sequence to see the day and the alternatives within each moment.
- Look more closely at potential keepers. It opens those photographs larger and compares nearby frames.
- Mark its suggestions Red in Photo Mechanic. These stay separate from my own T tags.
- Review the whole shoot myself. I compare our choices, add photographs, and look around the moments it has identified.
- Edit and refine in Lightroom. I make the final choices as the photographs and the gallery come together.
The useful part is that the suggestions appear somewhere familiar. Photo Mechanic remains my culling workspace, and Lightroom remains my editing workspace.
What the close calls taught us
Keeping our selections separate also gives us something concrete to discuss. We can see where we agree, where I’ve chosen another frame, and what tipped the decision.
In one veil sequence, the ChatGPT agent preferred a photograph where the bride’s arms were still raised. I preferred a nearby frame where her arms had dropped, but some movement remained in the veil.
The second photograph felt more natural. You could see less of the work that had gone into making it happen.
In a dance sequence, my preference came down to the symmetry of the dress and the pose. Both photographs had movement and energy. The difference was in how everything came together within the frame.
Some portrait decisions involved even smaller details: a more flattering expression, foliage touching or intersecting with someone’s head, or the couple’s noses meeting at an awkward angle.
Two frames can look almost identical until you put them beside each other. Then the difference becomes the only thing you can see.
Shoot more, less often
Some advice I heard once that stuck with me was to “shoot more, less often.” During something emotionally important, such as a first look, I might make a lot of very similar frames. I’m watching for tiny changes in expression, gesture and connection.
Sometimes I’m also shooting through an awkward moment, waiting for the couple to relax or for a more natural expression to appear. The repetition has a purpose.
That has been useful to explain to GPT. A sequence might contain several distinct emotional moments, or one particularly flattering expression among otherwise similar frames. I want those differences considered before useful alternatives disappear from the selection.
Over the years, I’ve noticed that I typically deliver around a hundred photographs per hour of coverage, including both my work and my second photographer’s. That’s an observation, rather than a target I’ve drawn up. Some weddings are more action-packed; others are quieter affairs. It gives the ChatGPT agent context, but the selection still needs to follow the wedding.
Showing it complete delivered galleries has helped too. Those galleries show how I use details, wider scenes, quiet moments and sequences within the finished story. They also give it a better sense of how much variety I might keep from one part of the day.
We looked at weddings in different settings and light, including darker portraits at Duke Chapel and brighter outdoor weddings. That mattered because I wanted the ChatGPT agent to understand the different moods I work towards.
Why a dark preview can still be a keeper
I often use a black-and-white preview in-camera because it helps me remove the distractions of colour and concentrate on light, shape and composition. The finished photograph may still be delivered in colour or, if I like both treatments, in both colour and black and white.
Exposure is another consideration. A preview may look dark because I’ve purposefully underexposed it with the final RAW edit in mind. In other situations, I may deliberately overexpose relative to how I want the finished photograph to look.
I’m thinking about the light, the detail I want to retain, and how I intend to process the file. The preview is one stage in that process.
The ChatGPT agent needs that context. A dark face or a bright preview doesn’t automatically make a photograph a reject. Equally, it can’t know exactly how far I can take a RAW file simply by looking at the preview. That is something I assess during review and editing.
A first pass while I slept
Once we’d worked through that first cull and compared our choices, I gave the ChatGPT agent a queue of four other weddings awaiting attention.
Then I went to bed.
While I slept, it worked through first passes of those weddings, marking its suggestions Red and preserving my existing tags and earlier selects.
I’ve done a rough review of the initial tags, and they look good. I haven’t yet gone through those four other weddings in full, so that’s the next stage of the experiment. I’ll review and fine-tune those selections myself, then update this post accordingly.
For now, the early results are promising. Coming back to a first pass already waiting in Photo Mechanic is certainly an appealing way to start.
Culling with ChatGPT: still a work in progress
There has been troubleshooting along the way, and the first wedding gave us plenty of close calls to discuss. Each time I preferred a different frame, I could explain why and add that guidance to the process. We’re building a set of examples and preferences for the ChatGPT agent to refer to.
I haven’t decided that this is permanently how I’ll cull every wedding. I may change the approach, use it for certain jobs, or return to doing the whole thing myself.
For now, it has given me a useful starting point and an interesting way to examine my own decisions. I’m sharing it because other photographers might find the experiment useful too—and because some of you may not have considered giving GPT this particular job.
There’s still gold to find. I’ve just got some help deciding where to start panning.
