Why AI Walk Videos Look Fake (and What Helps)

People say a clip "looks AI" and usually cannot say why. That vagueness is expensive: it makes people redraw the same bad photo five times, blame the tool, and leave.

The uncanny look has causes, and they are not equally fixable. Here are the four that actually show up, ordered by how much of the problem they explain.

1. The face is doing something the body is not

This is the biggest one. A walk has a rhythm: weight shifts to the front foot, the shoulders counter-rotate, the head stays almost level. If the face in your photo is smiling at a camera while the body performs a neutral walk, the two readings fight, and the eye notices the mismatch long before it can name it.

The fix is not a prompt. Use a photo where the expression matches the action: relaxed, looking at the lens or slightly past it. A candid frame beats a posed portrait for exactly this reason.

2. The light does not match between face and scene

A photo shot at noon under hard sun, dropped into a scene with soft evening light, produces a subject that looks pasted on. Graders of "looks fake" almost always include this, even when nothing is technically wrong with the render.

You cannot change the scene; the background comes from the reference. What you can do is pick a photo whose light is soft, directional and unremarkable: open shade, an overcast day, a window on the side. Which is why the photo checklist matters more than any instruction you write afterwards.

3. Edges: hair, glasses, jewellery

Fine detail is where generative models spend their error budget. Loose hair strands, thin wire glasses and reflective jewellery are the three things most likely to shimmer or weld to the skin between frames.

You do not need to fix this perfectly; you need to stop choosing the photo that maximises it. Hair tied back, no reflective accessories, no sunglasses. If a detail is important to the look, it is also the detail most likely to break.

4. Compression on the way in

A photo that has been through three messaging apps has block artefacts around the edges and smeared texture over the skin. The model reproduces what it receives, then the twelve second render compresses again, and the result reads as soft in a way the original never was.

Send the original file. Not a screenshot of the photo, not a forwarded copy: the file the camera produced.

What no prompt can fix

Two things, and it is worth knowing them before you spend credits.

The first is the camera itself. The move is copied from the reference clip, so it will have the reference's rhythm rather than yours. If you want a slow push-in, this tool will not give you one.

The second is lip sync. There is no dialogue in these clips; the model is not tracking phonemes. Any frame-by-frame expectation of speech will read as wrong.

What a prompt can do is narrower and real: it can nudge the look of the clip (wardrobe, setting mood, how the person holds themselves) without changing the choreography. Eight examples that actually move the needle are here.

One variable at a time

When a render comes back wrong, change exactly one thing before the next attempt. Swap the photo and keep everything else identical, or keep the photo and change the tier. Changing three things at once produces a clip you like and no idea why, which is the expensive version of learning.

A useful habit: keep a note of which photo, which tier and which one sentence you used for each attempt. After three attempts you will see the pattern in your own photos, and the fourth render will be the one you keep.

The ranked checklist

Before you render, score the photo honestly:

1. Expression matches a neutral walk (not a posed smile).
2. Light is soft and comes from one side.
3. One person, facing the camera, knees-up or full body.
4. Hair controlled, no reflective accessories, no sunglasses.
5. Original file, not a forwarded or screenshotted copy.

Four out of five is usually enough for a usable clip. Two out of five is a wasted credit, and the honest move is to take a new photo rather than re-roll the same one.

A note on expectations

A twelve-second vertical clip that copies someone else's camera move will never look like a shot you directed on a set. The bar is not "convincing for a film"; it is "convincing in a phone feed, for twelve seconds, with the sound on". Judged against that bar, most of the uncanny look disappears once the four causes above are handled.

If it does not, the photo is the thing to change. Open the studio and try the 10-credit tier, with a failed render refunded, and compare it against the photo you were about to blame the tool for.