What AI couple photos can and cannot do

What it does well

Placing two recognisable people into a scene that never happened is, at this point, largely solved. Lighting matches, skin tones sit correctly against the background, and the composition reads as a photograph rather than a collage. For a couple who simply never got round to a proper shoot, the result is often better than anything on their phones.

It is also very good at variation. Four takes come back from each generation with different lighting treatments, and the useful one is frequently not the one you would have predicted.

Where it still fails

Hands remain unreliable. Fingers interlocked between two people is one of the hardest things to render, and a proportion of results will have something subtly wrong — an extra knuckle, a thumb at an impossible angle. Poses where hands are visible and touching fail more often than poses where they are not.

Fine text does not survive. Signage, printed invitations, anything with small lettering in frame will come back as convincing-looking gibberish. This is a known limitation of image models generally, not of one product.

Very specific jewellery and tattoos are approximations. If a particular ring matters to you, expect the model to produce a ring, not that ring.

It is not a photograph of an event

This is the honest framing. A generated wedding photo is not evidence that a wedding happened; it is a picture of what the two of you would look like at one. People use these for save-the-dates, for previewing a dress silhouette or a venue style before booking, for a gift, and for the photo they never took. They are not documentation.

We say this plainly because the alternative — implying otherwise — leads to disappointed buyers, and a refund is more expensive to us than a clear expectation.

How to work with the limitations

Pick scenes where hands are not the focus. Prefer wider compositions over tight close-ups if you want a safe result, and tight close-ups if you want a striking one and are willing to regenerate. Generate more than once for anything you intend to print; the four variations exist precisely because the good one is not guaranteed on the first attempt.