Every image-to-video model claims to "preserve" your reference image. None of them preserve all of it. What they keep, and what they quietly redraw, is different per model — and knowing the difference is the whole skill.
We ran the same six reference frames through Veo 3.1, Kling 3.0 Pro, Wan 3.0 Prime and Seedance 2.5, with the same neutral prompt, and compared frame 1, frame 60 and frame 120 against the reference. This is what we found.
What holds
Composition. If your reference has a subject in the left third and a horizon at two-thirds height, every model we tested kept that arrangement through frame 120. Camera position survives too, as long as the prompt doesn't ask it to move.

Lighting direction. Key light from the left stays from the left. Colour temperature is less reliable — see below — but the direction of the light is one of the most stable things in the frame.
Large shapes. Buildings, mountains, the silhouette of a car. Anything that occupies more than about a tenth of the frame tends to keep its outline.
What drifts
Faces. Every model. Kling holds identity longest — around four seconds before the jaw starts to migrate — and Wan shortest. Hailuo-03's four-reference mode is the only thing we've found that meaningfully extends this, and it gets its own post.
Text and logos. Legible in frame 1, soup by frame 30. If a label matters, composite it in the editor.
Colour. Models pull colour toward whatever the prompt implies. A neutral grey reference with the word "dawn" in the prompt will be warm by frame 20. This is the most common cause of "it ignored my image" — it didn't, it obeyed the prompt.
The rule that works
Put the thing you can't afford to lose in the reference. Put the thing you want to change in the prompt. Never put the same thing in both.

The same scene, the camera pushes in slowly, lighting unchanged
The prompt above is deliberately empty of description. Everything about the look is in the reference; the prompt only says what should change. Renders from prompts written this way scored highest for fidelity on every model.
Don't describe the reference back to the model ("a woman in a red coat on a pier") — you're giving it permission to redraw her. And don't put a negative prompt in to protect the image; on four of five models it made drift slightly worse.
Per-model notes
- Veo 3.1 — best colour fidelity of the four; weakest at holding a face past three seconds.
- Kling 3.0 Pro — best identity hold; will sometimes add motion to a still subject you wanted static.
- Wan 3.0 Prime — fastest and cheapest; treats the reference as a strong suggestion rather than a constraint.
- Seedance 2.5 — closest frame-1 match of any model, then drifts fastest.
| Strength | Watch for | |
|---|---|---|
| Veo 3.1 | Best colour fidelity | Face past three seconds |
| Kling 3.0 Pro | Best identity hold | Adds still-subject motion |
| Wan 3.0 Prime | Fastest and cheapest | Strong suggestion |
| Seedance 2.5 | Closest frame-1 match | Drifts fastest |


