The best photos for AI video recasts (and the ones that fail)

A recast lives or dies on the input photo. The video model has to hold your likeness steady across hundreds of frames, from angles your photo never showed. Give it a strong reference and it has something to anchor to. Give it a weak one and it starts improvising.

What the model is actually doing

Reference-to-video models like Seedance take your photo as an identity reference. On every frame they try to draw a person who matches it: face shape, skin tone, hairline, build and clothing. The less ambiguous the reference, the less the model has to invent.

Photos that work

Front-facing, eyes visible. Most of what makes a face recognisable is between the eyebrows and the chin. A straight-on photo gives the model all of it.

Soft, even light. Window light or open shade. Hard shadows across the face get baked into the likeness and then drift as the scene's lighting changes.

Waist-up or full body. The model needs to know your build and outfit to put you convincingly in the scene. A tight face crop leaves the body to chance.

Just you. One person, uncluttered background.

Photos that fail

• Sunglasses, caps, masks. Hidden features get filled in, differently on every frame.
• Heavy filters and beauty modes. Smoothed skin and altered proportions give the model a face that doesn't quite exist.
• Group photos. The model may pick the wrong person or blend two.
• Low-resolution crops. A face that's 80 pixels wide doesn't carry enough detail.
• Extreme angles. Profiles and sharp top-down selfies lose the shape of the face.

Quick checklist

1. One person, facing the camera
2. Good light, no harsh shadows
3. Nothing covering the face
4. Upper body visible
5. Original photo, no beauty filter

If a render still comes out wrong, the credit for a failed video is refunded automatically. A different photo usually fixes it.