Why Can't AI Deform a Face Convincingly?

Updated 2026-10-27·3 min read·Rendering

What we tried, and how each one failed

We set out to make a real actor's face deform, like clay giving way. In one session it failed five ways:

• Words alone in a reference-to-video prompt on Wan gave a small textured patch on one cheek, not a changed face.
• An image model asked for "clay", "putty", "grey" or "glossy" produced something that read as makeup sitting on the skin. Our director's verdict: "makeup just laying on it."
• An image model told "it's his own skin, the anatomy moved" softened the face into a droopy but normal one. It wouldn't redraw grotesque anatomy.
• A code-made warp used as Wan's end frame didn't hold. Wan normalized it, and the take ended on a near-normal face.
• The same warp applied in post over the finished take used real pixels and didn't look like makeup. But a smooth geometric bend with no change in lighting read as a funhouse-mirror filter.

Why it happens

These models learn what faces look like from enormous numbers of ordinary photos. Push them toward an impossible face and they do one of two things. They keep the real face and lay the change on top (makeup), or they treat the change as a mistake and correct it back toward normal.

There was one partial success, and it explains the rest. When we gave Wan a start frame and an end frame where the end frame looked like a plausible photo (the clay-makeup version), it traveled to that end frame faithfully. So a video model will follow an end frame that looks possible and reject one that looks impossible.

What to try instead

1. Go small and wrong. Try an eye that has settled a little lower than the other. Something moving under the skin of the cheek. Subtle wrongness is far easier for a model to accept than a melt. We haven't proven this route yet, but every failure points toward it.
2. Make every keyframe a believable photograph. If you use start and end frames, the end frame has to pass as a real image, or the model will quietly fix it.
3. If you warp, repaint it. A geometric warp needs new folds, creases and shading painted in before it reads as flesh instead of a filter.
4. Let the edit do some of the work. A reaction shot, a cut away at the peak, or a shadow across the face can sell what the pixels can't.
5. Approve the stills before you render. Deformation takes several attempts, and each video render is the expensive part.

An example

For a scene where a man realizes his face is changing, don't ask for the whole face to sag. Start with a still where one eye sits a few millimeters low and the skin beneath it bulges slightly. Render a short close-up with that as the start frame, cut to the other character's reaction, then return to a frame that's a little worse. A few small steps like that read as wrong better than one impossible jump.

In ShowMaker

You can lock the actor's face with their character sheet, approve each still on the storyboard, and watch the whole sequence as a free animatic before paying for video. If an image model refuses an intense frame outright, see why AI video prompts get blocked.