AI filmmaking
How to Keep the Same Face Across AI Video Shots
Your character is a woman in a red coat. In shot one she is thirty. In shot two she is forty-five and the coat is maroon. Nothing you typed changed — and that is exactly the problem.
Why it happens
Every generation starts from noise. The model is not remembering your character between shots, because there is nothing to remember — each clip is an independent act of invention that happens to be guided by similar words. "A woman in a red coat" describes a region of possibility, and the model picks a different point in it every time.
Understanding that one fact changes how you work. You are not fixing a bug. You are narrowing a range.
The four methods, worst to best
1. Describing harder
The instinct is to add adjectives: brown eyes, shoulder-length dark hair, late thirties, small scar above the left eyebrow. This helps, and it is where everyone starts, but it plateaus fast. Twenty words of description still leave millions of faces, and past a certain length the model starts dropping details from the end of your prompt.
Worth doing as a baseline. Not worth relying on.
2. A reference image
Most current tools accept an image as well as text. Give them a picture of your character and consistency improves sharply, because you have replaced a description with an example. This is the single biggest improvement available to most people, and it is often the only one needed for a short film.
The catch is that the reference has to be the same image every time. Generating a fresh "reference" per shot reintroduces the drift you were trying to remove.
3. Seeds
Where a tool exposes a seed, the same seed with the same prompt gives you the same output. Changing the prompt even slightly while keeping the seed gives you something related. This is powerful for variations on one shot and close to useless across different shots, because the moment the prompt changes meaningfully the seed stops anchoring the face.
4. Designing the drift out
The method nobody mentions, and the most reliable one: write scenes that do not require a face to persist. Back of the head. Hands. Reflections. Over-the-shoulder from behind. Silhouette against a window. A character can carry a whole scene without a consistent face if the shots are chosen to make that true.
This is not a compromise or a workaround. Directors have used exactly these shots for a century, for reasons that have nothing to do with AI.
What drifts first
Not everything degrades at the same rate, which matters because you can spend your effort where it counts:
| Feature | How fast it drifts | What to do |
|---|---|---|
| Exact facial structure | Immediately | Reference image, or avoid the face |
| Age | Fast | State it as a number, not "young" |
| Hair length and colour | Fast | Repeat it verbatim in every prompt |
| Clothing colour | Medium | Name the exact shade, not the colour |
| Clothing shape | Slow | Usually holds on description alone |
| Build and height | Slow | Usually holds on description alone |
The pattern: the more specific a feature is to a particular human being, the faster it goes. General attributes survive; identity does not.
"Red coat" will drift across every red there is. "A dark red wool coat" holds far better, because you have replaced a category with a point. Do the same for hair, light and location — specificity costs nothing and buys consistency.
The copy-paste block
Keep your character as a fixed block of text and paste it into every prompt unchanged. Not rewritten from memory, not paraphrased — pasted. Something like:
A woman in her late thirties, shoulder-length dark brown hair, dark red wool coat, plain grey scarf.
Then the shot-specific part goes after it. The discipline of never retyping that block is more valuable than any individual word in it, because the drift you are fighting is as often yours as the model's.
When to stop
If you have spent an hour and six generations on one shot of a face, the scene is asking for something the tool cannot give you today. Rewrite the shot rather than keep paying for it. A cut to her hands, or to what she is looking at, almost always works — and often plays better than the shot you were trying to force.
This is the part people resist, and it is the part that separates a finished film from a folder of near-misses.