Why Faces Change in AI Videos - and How to Keep Them Consistent

The short version

A face changes in an AI video because the model is not pasting your photo into the frame - it is redrawing that person in every frame, using your photo only as a reference. Anything that makes the reference harder to read, or that forces the model to guess, shows up as drift: the eyes widen slightly, the jaw shifts, and by the end of a ten-second clip the person is almost but not quite themselves.

The fix is not a better model or a longer prompt. It is the input: one clear, front-facing photo per performer, with no second face anywhere in the frame.

Why a face changes at all

Video models generate frames, not copies. Each frame is a fresh reconstruction of what the model believes the scene looks like, and identity is one of the things it has to reconstruct. A face is a hard thing to hold steady, because so much of what makes someone recognisable is small: the exact spacing of the eyes, the shape of the jaw, an asymmetry most people never notice.

That is why the failure looks the way it does. You rarely get a completely different person - you get a sibling. The model kept the broad features it was confident about and invented the details it was not.

Three things decide how much it has to invent: how much pixel detail the reference gives it, how many people it has to keep apart, and how much of the scene is moving at the same time.

The three causes behind most bad renders

Almost every case of drifting faces traces back to one of these inputs:

• Two people in one photo - the model has to decide which face belongs to which performer, and usually blends them
• A face that is small in frame, blurred, or in profile - there are too few pixels to reconstruct from
• Heavy beauty filters - smoothing removes exactly the small asymmetries that keep two identities apart

Why this format is unusually stable

The two-photo booth format sidesteps most of this, and that is by design rather than luck.

The scene never changes. The backdrop, the hanging microphone, the lighting and the camera move come from a fixed plate, so the model does not spend its capacity inventing a room. Everything it knows about the frame is already settled, which leaves identity as the main variable it has to solve.

The roles are explicit. Two separate photos go in, and the first is assigned to the left performer and the second to the right. There is no ambiguity about who is who, which is the single most common cause of blended faces. You can read the exact wording on the hotel lobby ai prompts page - it is published in full.

The camera is locked. A fixed medium-wide shot with no cuts means no new angles to re-solve. Clips that cut between shots are where identity drifts fastest, because every cut is a fresh reconstruction.

The photo checklist that prevents it

Run both photos against this list before you upload. It takes about a minute and it is the difference between a first try that works and three that do not:

• One subject per photo - if there is a second face in the background, crop it out or pick another photo
• Front-facing, eyes visible, no sunglasses and no strong side profile
• Even lighting on the face; backlit and flash-blown photos lose detail in the shadows
• Reasonably sharp - a slightly soft photo is the most common cause of unstable results
• Similar crop and framing for both people; portrait and square both work
• No heavy beauty filters or skin smoothing

If it still drifts

Check the frame edges of both photos first - a bystander half out of shot is still a second face to the model. Then check the order: the first photo becomes the left performer and the second becomes the right, so a pairing that reads wrong may just be uploaded in the wrong order.

If the two people look very similar, or the photos are lit very differently, generate one test before committing to a larger pack. A render only consumes a video when it completes - provider errors, timeouts and content-safety refusals restore it to your balance automatically.

The rest of the preparation is covered step by step on the how to make a hotel lobby ai video page, and the shipped pairings are collected on the hotel lobby ai template page.

Turn it into a clip

Open the hotel lobby ai video generator, upload one photo per performer, and press generate. The scene and camera move are applied for you, and the clip arrives silent; there is no prompt to write.

Every render is a ten-second 720p widescreen clip. Packs are one-time and start at $4.90 per video - see the pricing page for what each pack includes, and there is no subscription.