How to make two people kiss with AI

You give the app two faces and it gives you back a short video of those two people kissing. There is no timeline, no editing and no prompt to write: choose photos, choose a frame, choose a style, press generate, download an MP4.

The part worth understanding is the middle. AI Kiss does not animate your two photos separately and stitch them together. It first builds one single image containing both people, and then animates that image. Once you know that, most of the advice about which photos to use stops being arbitrary and starts making sense.

When you add a photo, the app reads it on your device rather than uploading it immediately. It fixes the rotation, because phone photos carry orientation data that would otherwise turn a portrait sideways, and it scales the image down so the longest edge is at most about 1600 pixels. That is plenty of detail for a 720p clip and it keeps the upload small.

Then it composes the keyframe on a canvas. In two-photo mode the frame is split down the middle and each photo is fitted into its half, cropped to fill rather than squashed. In one-photo mode your picture fills the whole frame. The result is one JPEG: two people side by side in a single image, which is the scene the model is asked to animate. That composed image is the only thing uploaded, and both of your originals stay on your device.

The keyframe is sent to the generation service with the style you picked, and it is queued. The model treats the keyframe as the first frame of a scene and predicts what happens next: heads turning, the distance closing, a kiss, a small settle. The style you chose is the instruction that shapes that motion, so Kiss produces a normal closed-mouth kiss, French kiss a deeper and slower one, and Hug an embrace instead.

Everything visible in the clip comes from your keyframe. The faces, the clothes, the backgrounds behind each person. The model is not fetching a stock couple and pasting your faces onto it; it is moving the pixels it was given. This is why two photos with matching light look like one scene and two mismatched photos look like a collage that has learned to move.

While it runs, the app checks the job every few seconds and shows the status on the My videos screen. You do not have to keep staring at it, but you should come back in the same browser, because your library is tied to that anonymous session.

An MP4 of about five seconds, at the aspect ratio you chose, with no watermark. Play it in the browser, download it to your camera roll or desktop, or share it straight out of the app.

Each generation costs one credit, bought once, with no subscription and no account. The pricing page has the current figure. If a generation fails, the credit returns to your balance automatically, so you are not charged for a broken render and you can retry from the My videos screen with the same photos or better ones.

There is no edit step afterwards. If you want a different style or a different framing, that is a new generation. Real finished clips are on the examples page if you want to see the output quality before spending anything.

Sharpness first. The model animates the detail it is given and cannot invent what the camera never captured, so a small sharp photo beats a large blurry one every time. Motion blur, noise and heavy beauty filters all cost you.

Angle second. A face pointing at the camera, or turned at most a quarter away, gives the model both sides of the face. A strict side profile forces it to guess at the half it cannot see, and guessed faces are the ones that look wrong. Eyes and mouth should be visible: no masks, no hands across the face, no hats shading the eyes.

Consistency third, and this is the one people skip. The two photos have to belong in the same picture: similar brightness, similar distance from the camera, similar image quality. If you already have a photo of both people together, use one-photo mode. The photo requirements guide has the full checklist, and how to create AI kiss videos walks through a first attempt.

Three failure modes cover almost everything. Blurry or low-resolution faces produce blurry, smeared motion. Side profiles produce distorted features as the model invents the far side of the head. Group photos with three or more people leave it ambiguous who should kiss whom, so crop to two faces before uploading.

The fourth is subtler: mismatched sources. A crisp studio portrait next to a dim phone snapshot never merges into one believable scene, and neither does a black and white photo paired with a bright colour one, which is the common trap with old family pictures.

There are also things the tool does not do: no audio, no clip longer than about five seconds, no editing after the render, no control over the background. If a generation fails outright, the refund is automatic and the retry costs nothing extra.

Use your own photos, and photos of adults who have agreed to appear in the video. No photos of minors, ever. No celebrities or other public figures without their consent, and no pictures of strangers pulled from social media. This is in the terms.

The composed keyframe and the finished video are stored so the app can show you your library, and you can ask for them to be deleted by contacting support, which the privacy page explains. If the clip is going anywhere public, tell the other person first and say in the caption that it is AI. The privacy and safety guide goes deeper.