Every few months, the image-generation models get a notch faster, a notch cheaper, and a notch more consistent. Individually each release is a footnote. Stacked up, they quietly move the line on something very practical: what it costs a D2C brand to photograph a catalog.
The short version
Studio-grade product imagery used to mean booking a studio, a photographer, models, and a week of turnaround. A growing share of that can now start from flat-lays and a prompt — and the part that's automatable keeps getting cheaper faster than the part that isn't.
What actually changed
- Speed — generations that took a minute now take seconds, so iterating on a look is cheap.
- Cost — per-image cost has fallen far enough that generating dozens of variants per SKU is reasonable.
- Consistency — newer models hold a product's shape, colour, and texture across scenes far better than the early ones did.
A simplified version of the pipeline we run looks like this:
# flat-lay in → lifestyle set out
catalog = load_skus("./products")
for sku in catalog:
base = segment(sku.flatlay) # isolate the product
scenes = render(base, style="editorial", n=6)
keep = qa(scenes, brand=sku.brand) # human picks the winners
publish(keep)Where it helps a D2C catalog
- Hero shots for new drops without a studio booking.
- Variant coverage — every colourway and size shown, not just the ones you had budget to shoot.
- Lifestyle & seasonal sets — same product, many contexts, on demand.
- Marketplace compliance — clean, consistent backgrounds across hundreds of listings.
The win isn't "free photos." It's being able to show every product, in every context, without that being a budget decision.
Where it still needs a human
Faster models don't remove judgement — they relocate it. The things that still go wrong without a person in the loop:
- Brand consistency — a model will happily drift off your palette and styling unless someone's steering.
- Fine detail — fabric seams, logos, small text, and hands are still where artefacts hide.
- Truthfulness — the product shown has to be the product shipped. Generated imagery that misrepresents an item is a returns-and-trust problem, not a photography one.
How we use it at Codeja
This is exactly the workflow behind our AI Product Photography service: outfit and product flat-lays in, a reviewed lookbook out, with a human approving every frame against the brand. The faster models just mean we can offer more variants per SKU for the same turnaround.
If you're staring at a catalog you can't afford to shoot the old way, send us a few SKUs and we'll show you what comes back.