Why your AI kiss video doesn't look like you

Your two photos are combined into one image before anything is generated, and that single image is all the model ever sees of you. It is not looking anyone up, and it has no second reference to check itself against. Every frame after the first is predicted from those pixels, which is why the resolution, sharpness and angle of the source photo set a hard ceiling on how much you can recognise the result.

That ceiling gets lower as the clip goes on. Image-to-video models hold a face together confidently for a few seconds and then start to drift, so the beginning of a clip typically looks more like you than the end of it. If likeness is the thing you care about most, a 5-second clip is the safer choice than a 15-second one, even though the longer option exists.

The other quiet destroyer is compression. A photo that has been sent through a messaging app, screenshotted, re-sent and saved again has lost exactly the fine detail around the eyes and mouth that makes a face read as a specific person. Go back to the original file if you have it. The full flow is on the how it works page.

This is not a problem with one tool. When we read 1,043 one- and two-star App Store reviews across 52 AI photo and video apps on 13 September 2026, "it looks nothing like me" was the widest-spread quality complaint in the entire corpus: 123 reviews, across eighteen separate apps in the kiss and hug niche alone. The quotes below are verbatim, as complaints made on a date rather than claims about what any app does today.

"Very limited and when you use photos of yourself the video issued looks nothing like you" · "Looks nothing like me", which was one reviewer's entire review · "It just changes stuff just a bit but it looks nothing like us" · "You'll never look like yourself when the result comes out." · "the face is all distorted" · "After that, it started doing weird stuff that didn't look like me."

Two things are worth noticing about that list. It spans expensive tools and cheap ones, big apps and tiny ones, which is the tell that the cause is not the brand. And almost none of the reviews mention what was uploaded, which is the one variable that would have explained most of them.

In rough order of how much difference they make, and all five are decided before you press generate.

The same handful of source-photo problems produce the same handful of complaints. This is the lookup table.

No accuracy guarantee, from us or from anyone selling this kind of tool honestly. Image-to-video models drift, hands and arms are hard, and the same pair of photographs can produce a clip you love and a clip you delete. Anyone promising a guaranteed likeness is describing something the technology does not currently do.

What is actually guaranteed here is narrower and checkable. If a generation fails — an error, a job that never completes — every credit it cost returns to your balance automatically, for any length, without you having to ask. That covers failure, not disappointment: a clip that generates successfully but is not flattering has been generated, and the credits for it are spent, the same as on every other tool.

Which is why the examples page shows the two ordinary photographs underneath each clip. You can see the input quality that produced each result before you spend anything, rather than judging the tool on a marketing render made from a studio portrait.

Change one thing, not everything. The usual winning change is a different source photo of the same person: sharper, more front-on, better lit. People tend to reach for a different style or a different aspect ratio first, and neither of those touches the likeness.

If both photos are already the best you have, drop to a 5-second clip. Less time on screen means less room for drift, and the opening seconds are consistently the strongest part of any generation.

And if the photo you want to use is a scan of an old print, scan it again rather than photographing it: flat, in good light, no glare across the face, at the highest resolution your scanner offers. Old photos are the most common source of both the best and the worst results on this site, and the difference is almost always the scan.