The best photos for AI video are almost never the best photos on your phone. They are the ones that encode cleanly: a large, sharply lit face with strong edges and no filter. Everything else — composition, mood, wardrobe — is decoration the model will overwrite anyway.
Five criteria predict the result well enough that you can grade a photo in about ten seconds. They are listed below in the order that matters.
One: face size beats everything
A model reconstructs identity from the encoded face, and that encoding has a resolution ceiling. Below roughly 300 pixels across the face, distinguishing features start to average out and the output drifts toward a generic person.
| Shot type | Face width at 1080px | Verdict |
|---|---|---|
| Headshot | 500-700px | Best |
| Waist-up | 250-350px | Workable |
| Full-body | 100-150px | Usually fails |
| Group of four | 80-120px each | Always fails |
If you only have a full-body shot, crop it. A crop loses the outfit and keeps the face, and the face is the only part the model cannot invent.
Two: hard light, one direction
Soft, even light is flattering to humans and useless to models. The model needs contrast to locate the edges of a nose, a jaw, a cheekbone. A single hard source — window light from one side, a phone torch, a street lamp — produces a face with structure.
Overhead bathroom lighting is the single worst common case. It flattens the eye sockets and produces a waxy, mannequin-like result that people blame on the AI.
Three: a neutral expression that still has a mouth line
Closed-mouth, relaxed, eyes open. Two failure modes:
- Wide open smile with teeth — the mouth is the hardest region to keep coherent across frames, and teeth are where artefacts appear first.
- Deadpan stare — encodes fine, but the render inherits the flatness and the clip feels lifeless.
The middle ground is a small, closed smile. It gives the model a lip line to hold and the viewer something human to read.
Four: angle, not profile
Three-quarter angle — face turned roughly 45 degrees from the camera — is the strongest input for most systems. It gives the model depth information that a flat frontal shot lacks, and it still keeps both eyes visible.
Pure profile is tempting because it looks cinematic. It is also the angle where identity is least constrained, so the model has the most freedom to drift.
Five: nothing that has to be invented
Anything the model cannot see clearly, it will replace. That includes:
- sunglasses or heavy shadow over the eyes
- hands near the face
- hair covering one side of the jaw
- busy backgrounds with high-frequency texture
- novelty filters, heavy grain, or beauty smoothing
Google's guidance on image quality for search is written for a different problem but arrives at the same place: sharpness and a clear subject beat resolution. The render agrees.
Why these five criteria and not composition
Because the model is not looking at your photo the way a person does. It encodes a compressed summary, then generates new frames conditioned on that summary — the mechanism described in the Stable Video Diffusion paper as inserting temporal layers into a still-image model and fine-tuning on video.
"Latent diffusion models trained for 2D image synthesis have been turned into generative video models by inserting temporal layers and finetuning them on small, high-quality video datasets."
The consequence for photo selection is direct: what survives compression is structure and contrast. What does not survive is everything you would normally call "a nice photo".
A ten-second checklist for the best photos for AI video
Run any candidate photo through this and score it:
- Can I see both eyes clearly? (no → reject)
- Is the face at least a third of the frame height? (no → crop or reject)
- Is there one obvious light direction? (no → reject)
- Is the mouth closed or slightly smiling? (no → expect artefacts)
- Is the background plain? (no → crop tighter)
Three or more yes answers is enough to try. Five is a photo worth keeping for future renders.
Grading also saves money. A render costs 8 credits, and the $7.99 pack holds 10 — which means one careless upload consumes almost the whole pack. Spending ten seconds on the checklist before you spend 8 credits is the highest-return habit in this workflow.
Frequently asked questions
Are phone photos good enough as source images?
Yes, if the light is directional and the face is large in frame. A modern phone in a doorway with daylight behind you outperforms a DSLR shot in flat indoor light — the best photos for AI video are decided by light geometry, not by the camera's price.
Should I use the highest resolution file I have?
Not necessarily. Resolution helps only up to the point where the face is sharply resolved. A 12-megapixel photo with motion blur loses to a 4-megapixel photo that is crisp — and for the best photos for AI video, a 5-second clip rendered from the crisp one will look better in every frame.
Does the background matter at all?
Only if it competes with the face. Plain walls, sky and dark interiors all work; a patterned rug or a crowded street can leak into the generated frames.
Can I fix a bad photo with a better prompt?
No. Prompts change the scene, not the encoded identity. The photo is the ceiling.
Grade five photos you already own against the checklist, then render the best one twice with different motion. Open the generator and let the results settle the question.