Image-to-Video Prompting: What Changes When You Start From a Still

Published October 2026 · 11 min read

The most common mistake in image-to-video work is reusing a text-to-video prompt. It feels logical — the same model, the same vocabulary, the same six-slot structure. But the task has changed underneath you, and a prompt that describes a scene in full will actively fight the image you supplied.

The reason is simple. In text-to-video, the model has to invent everything: subject, appearance, setting, light. In image-to-video, the first frame already contains all of that. Your prompt is no longer a description of a scene. It is a description of change.

The one rule that fixes most image-to-video prompts

Stop describing what is visible. Start describing what should be different one second later.

Compare these two prompts for the same still photograph of a woman standing in a doorway:

Wrong (carried over from text-to-video):
a woman in a red coat standing in a wooden doorway, overcast light, medium shot, film grain

Right (image-to-video):
she turns her head slowly toward the camera, coat hem lifting slightly in the wind, hair moving a beat behind

The first prompt spends its entire budget telling the model things it can already see. Worse, it invites the model to re-render the scene from scratch rather than animate the frame it was given — which is exactly how you get a subject whose face subtly changes shape. The second prompt describes only motion, which is the only thing that actually needs to be generated.

What to put in the prompt instead

A workable image-to-video prompt has four parts, in this order:

  1. The primary motion — the one thing that must happen. “She turns her head.”
  2. Secondary motion — what follows from it. “Hair lags a beat behind.”
  3. Environmental motion — what else in the frame is alive. “Coat hem lifting in the wind.”
  4. Camera behaviour — usually “static” or one slow move.

Notice that appearance, lighting and texture are absent. They are already fixed by the source image. If you restate them, you are at best wasting tokens and at worst giving the model permission to change them.

The one exception is when you want to change something from the still — a colour shift, a light turning on, weather arriving. In that case describe the change explicitly, not the resulting state: “the window light dims as clouds pass” rather than “dim overcast light”.

Morphing: the three causes and their fixes

Morphing — the subject slowly melting into something else, faces drifting, objects changing identity — is the signature failure of image-to-video. It has three distinct causes, and they need different fixes.

Cause 1: too much motion requested

Ask for a full-body turn, a walk across the room and a head turn in five seconds and the model has to invent an enormous amount of new visual information. Where it lacks data, it interpolates, and interpolation is what morphing looks like.

Fix: request less. A head turn plus hair movement is enough for a five-second shot. If you need a walk, generate the walk as a separate shot and cut.

Cause 2: the subject is small in frame

When a person occupies 15% of the frame, the model has very few pixels to work with, and small errors compound across frames. Faces in wide shots are the worst offender.

Fix: crop the source image tighter before generating, or accept that the shot must be a wide establishing shot with no facial detail — and keep it short.

Cause 3: contradictory motion

“She walks forward while turning to look back over her shoulder while the camera pushes in” is three motions competing for one body. The model resolves the conflict by blending them.

Fix: one primary motion per shot. Everything else is secondary and small. If the shot needs three things to happen, it is three shots.

Choosing a good first frame

Image-to-video quality is decided largely before you write a single word. The source still determines what is possible.

Motion amplitude: small beats large

There is a strong inverse relationship between how much motion you request and how believable the result is. A two-second shot of someone blinking and shifting their weight is almost always clean. A two-second shot of someone sprinting is almost always compromised.

This is not a limitation you can prompt your way around — it is a property of how these models estimate motion. The practical consequence is that image-to-video rewards restraint. The best results come from asking for the smallest motion that serves the story:

Model behaviour differs more here than in text-to-video

Image-to-video exposes model differences more sharply than text-to-video does, because each model has a different policy for how strongly it honours the source frame.

If a shot keeps failing on one model, it is often worth re-running the same source image and prompt on a second model before rewriting anything. The difference in source-fidelity policy can be larger than the difference your prompt edits would make.

A complete example

Source still: a man in a workshop, mid-reach toward a tool on a bench, warm window light from the left, medium shot.

he completes the reach and picks up the tool, fingers closing around the handle, dust motes drifting in the window light, slight camera drift handheld

Four elements, no appearance description, one primary motion, one environmental detail, one camera note. This is the entire prompt. Anything more is likely to reduce fidelity rather than improve it.

When text-to-video is the better choice

Image-to-video is not automatically better just because it gives you more control. It has a real cost: you must produce or source a good still first, and that still constrains everything that follows.

Text-to-video is the better tool when you need the model to invent a world — an environment you cannot photograph, a lighting condition you cannot stage, a subject that does not exist. It is also better when you need a specific camera move that starts from a position you cannot capture as a still, and when you are iterating on composition rather than animating a fixed frame.

The practical rule: if you know exactly what the first frame should look like, use image-to-video. If you are still deciding, use text-to-video until you find the frame, then switch. Many workflows do both — generate stills until one is right, then animate it. That path costs slightly more but gives you a controlled starting point, which is often worth it for hero shots, and it avoids the common trap of animating a mediocre still simply because generating it took effort.

Checklist before you generate

If you want ready-made motion phrasing rather than writing it from scratch, the image-to-video prompt library has presets grouped by motion type, and the camera move reference covers which camera behaviours survive the process intact.

One closing note: image-to-video is generally more expensive per usable second than text-to-video, not because the per-second rate is higher, but because a failed generation wastes a good source frame as well as the credits. The cost breakdown covers how retries affect the total, and it applies here with more force than usual.