AI Product Photography Without a Photo Shoot
A real store had been using the same manufacturer photos since 2014. No photo shoot, no extra subscription, and one rule the images were not allowed to break.
The short answer
You can replace catalog product photos with AI images without booking a photo shoot. Point Codex at a folder of your existing product photos, give it your brand direction and the business problem rather than a prompt for pretty pictures, and it will plan the work, decide the right treatment per product, generate candidates, and organise the approved files. I did it for a real store using Codex alone, with no other image tool and no second subscription.
At a glance
- Time
- Hours, not weeks
- Cost
- No extra tools
- You need
- Codex + your photos
- Where
- Codex desktop app
This store had been using the same product photos since 2014. Flat background, straight-on angle, the same shot every other store selling the same stuff is using. That's normal, because you're either stuck with whatever the manufacturer hands you or you pay for your own shoot. Shoots run into thousands of dollars and weeks of waiting, and the moment a new product lands you start again.
One rule governed the whole project, and everything below follows from it. As I put it while I was working: "the image can look better, but it cannot lie about the product".
That single line is what separates this from generating nice-looking pictures. Shape, coverage, scale, colour, logo placement, and whether the item is shown in a context that makes sense. If a prettier image misrepresents what arrives in the box, it fails, no matter how good it looks.
Key takeaways
- 01
An image may look better; it may not lie about the product. Every candidate got checked against the real thing for shape, coverage, scale, colour and logo placement. Prettier is worthless if the customer opens the box and finds something else.
- 02
Give it the business problem and your brand direction, not a prompt. I handed over the brand playbook we'd built earlier: audience, tone, positioning, visual guardrails. Without that you get random AI images that don't belong to the same business.
- 03
Clear direction, then room to decide. I didn't script every step. I gave the goal, the brand direction, the product constraints and feedback when needed. It handled the workflow. That balance is what made it work.
- 04
Not every product wants the same treatment. Some read fine as a clean flat lay. Others had to be shown on a model, because the shape or length would mislead lying flat. Choosing per product was the best decision it made.
- 05
Codex acted as a production assistant, not an image generator. It planned, organised files, wrote the prompts, generated, reviewed, rejected, renamed for SEO and prepared everything for the site. The images were almost the smaller half of the job.
I stayed in the loop the whole way through, and that was a deliberate choice rather than a limitation.
I could have built this to run start to finish on its own: pull the images down, sort them, generate replacements, upload them back, never look again. That is possible.
I didn't, because this is client work. If it were my own store I'd have handed it off and moved on. This was going out under someone else's name, so I wanted to see exactly what was leaving my hands before it left. Where you land on that depends on what you're building and who it's for.
Prefer to watch
9:36The workflow, start to finish
Get your existing photos out of your store
I had no access to the photos outside the website, so I pulled them down from it. That store runs WordPress and WooCommerce, so I went into the file manager, archived the uploads folder, and downloaded the lot.
There are plenty of ways to do this depending on your platform. This is just how I did mine.
Point Codex at the folder
I downloaded the Codex desktop app and pointed it at the folder holding those photos.
Codex isn't a chat box where you ask something and get an answer. You give it a task and it goes into your files and does the work: opens things, changes things, tries things, then comes back and shows you what it did. That is exactly what a folder of a few thousand photos needs.
Explain the business problem, not the picture you want
I did not ask for pretty product photos and hope.
I explained the actual problem: the photos were accurate but boring. They worked as catalog shots, but they didn't create interest or help anyone imagine the product in use. I wanted better website images, plus a plan for social content.
Hand over the brand playbook
Then I gave it the master brand playbook we'd built in an earlier project: brand direction, audience, tone, and visual guardrails.
This step is what stops you getting random AI images. I needed everything generated to feel like it belonged to the same business and was made for the right customer.
Narrow to the products you actually sell
It asked for the source files, then helped work out which products were still active and which should be ignored, so it wasn't generating imagery for things that had stopped being sold. There were quite a few of those.
The product list got cleaned up before any image work started.
Split the work into two tracks
Website images and social images are not the same job.
Website images appear on product pages, so they have to stay accurate. Social images can be more lifestyle focused, but they still have to match the brand direction. Separating them meant each track could be judged against the right standard.
Test a style before you batch anything
It decided not to refresh everything at once, and started with tests instead.
The point of a test is to find a style that makes the images more useful without making the product inaccurate. Get that wrong at batch scale and you have a few thousand wrong images instead of five.
Run the batch, then review it as a batch
Once the test direction worked, it selected the next group of products, decided what kind of image each needed, wrote the prompts, generated the images, saved them into a candidate folder, and named the files properly.
Then it built a contact sheet. Rather than reviewing every image one at a time, I could see the whole batch together, which is far better for spotting anything inconsistent, repetitive or wrong.
What the AI actually decided
Choosing the treatment per product
Some products worked as a clean flat lay. Others needed to be shown on a model, because the shape or the length would be misleading laid flat.
It wasn't applying one style to everything. It picked the treatment based on what would make each product clearer and more accurate, and that was the most impressive decision in the project.
The rejection, and what happened next
One image looked good, but the sides of the product were completely wrong.
I flagged it and treated it as a rejection. It moved the incorrect image out of the candidate set and regenerated it, staying closer to the source photo. The loop became: create, review, reject if the product truth is wrong, regenerate, approve only when it passes.
It never assumed a generated image was usable.
The contact sheet
After each batch it assembled a contact sheet of everything it had made.
Reviewing images one by one, you compare each against the product. Reviewing them together, you see whether the set holds up: repetition, drift, one image that doesn't match the others. Different problems, and only the second view catches them.
SEO file names, not IMG_4471
Files were saved with names that made sense for the website rather than whatever the generator produced. Organised, consistent, and ready to upload rather than needing renaming later.
Small thing, and it's the sort of small thing that costs an afternoon when nobody does it.
The tracking sheet
Approved images were copied into a final approved folder and logged in a tracking sheet: project name, file name, suggested alt text, where on the site the image should be used, approval status, and notes.
That's the part that makes this a workflow rather than a pile of pictures. The alt text alone is a job most stores never get round to.
Sources
Questions people actually ask
Can AI really replace a product photo shoot?+
For catalog-style product images, yes, and this store's had been unchanged since 2014. What it cannot do is invent a product it has never seen. It works from your existing photos, so you need those first.
What tool did you use for AI product photography?+
Codex, from OpenAI, and nothing else. No separate image platform and no second subscription, which matters when you're already paying for tools.
Will the AI images be accurate enough to sell from?+
Only if you enforce it. The rule on this project was that an image can look better but cannot lie about the product, and every candidate was checked for shape, coverage, scale, colour and logo placement. Anything that misrepresented the product got rejected and regenerated.
Do I need to be technical to do this?+
You need to get your photos into a folder, which on WordPress means the file manager. After that you're describing your business and judging images, which is not a technical skill. It is a knowing-your-product skill.
How do I stop AI images looking generic?+
Give it your brand direction before it generates anything. I handed over a playbook with the audience, the tone, the positioning and the visual guardrails. Skip that and you get images that could belong to any store.
Should I let it run the whole thing automatically?+
You can, and I chose not to. This was client work going out under someone else's name, so I watched every step. On my own store I'd have handed it off. It depends on what you're building and who it's for.
What do I do when an image comes back wrong?+
Say so, and be specific about what's wrong with the product rather than what you dislike about the picture. Mine got the sides of a product wrong; flagging that moved it out of the candidate set and regenerated it closer to the source.
How long does this take compared to a photo shoot?+
A shoot is thousands of dollars and weeks of waiting, and it starts over every time you add a product. This ran in a working session, and the batch approach means new products join the existing process instead of triggering a new project.
About the author
Hey, I'm Traci
Digital Marketing Strategist
With over 25 years in this industry I help small business owners and creators turn their online presence into profit.
Keep going
The setup, the templates, and what I build next
Not more reading. The Claude setup guide you can follow tonight, and the templates I actually use. The page opens the moment you sign up.
Get the guideFree · you get it straight away