Image to video: what a still can become, and choosing a first frame that survives

Image to video takes a still and continues it. It does not know what is behind the subject, what the far side of an object looks like, or what happens outside the frame — and where the motion requires any of that, it makes it up.

This single fact predicts almost every failure. Drift, morphing, objects that change shape, backgrounds that dissolve — all of them are the model inventing where it had no information.

So the craft is largely in choosing a source frame and a movement that keep the model inside what it can see.

A source frame with deep shadow, motion blur or shallow focus contains ambiguity, and every ambiguous region is somewhere the model will invent during the move.

What survives well: sharp throughout, evenly lit, subject clearly separated from background, nothing cut off at the frame edge in a way that begs to be revealed.

What does not: heavy bokeh, blown highlights, complex overlapping objects, anything where you cannot tell what is in front of what.

If you are generating the still yourself, generate it for this purpose. A frame optimised as a still and a frame optimised as a video source are different pictures.

1. No movement, ambient only. Subtle life — cloth settling, light shifting. Almost always works.
2. Slow push in. Reveals nothing outside the frame and no new surfaces. Very reliable.
3. Slow pull out. Reveals the edges, which the model invents. Usually fine, occasionally strange.
4. Pan or tilt. Reveals a whole side of the scene. Expect invention.
5. Subject rotation or turning. Requires a surface never shown. Expect failure.

Design sequences from the top of that list. A piece built from pushes and ambient motion looks controlled; one built from rotations looks like a series of near misses.

Every frame is generated from the previous one's understanding, and small errors compound. The first second is usually excellent; the fifth may not be.

Generate shorter than you think you need and use the beginning. Where you need a longer shot, generate several and cut between them, or extend from a late frame as a new source rather than asking for one long run.

Always check the final frame before watching the clip. If it has drifted, the clip is unusable regardless of how good the opening was — and in motion you will not notice.

A common mistake is writing a full image description as the prompt for image-to-video. The still already carries all of that, and re-describing it invites the model to reinterpret what it can see.

Prompt the change instead: what moves, how fast, and what the camera does. Plus the constraint block — what must stay identical.

The camera pushes in very slowly. The subject stays still.
The product keeps its exact shape, colour and label position.
No cuts. No additional objects enter the frame.

Short, about motion, with the preservation stated explicitly.

A still where the subject is cut off mid-object, where several elements overlap confusingly, or where the interesting content is in a heavily blurred region will not animate well no matter how it is prompted.

Neither will a frame that requires the subject to do something — a specific gesture, an expression change, an interaction. These need the movement to be generated from scratch, which is a different and much harder problem.

In both cases the answer is a different source frame, chosen or generated for the movement you want. That is a two-minute change and it is more effective than any number of prompt attempts.