Image-to-video is not a single workflow. Sometimes you need the video to begin from a specific still. Sometimes the ending composition matters just as much. In other projects, you care less about the exact first frame and more about keeping a person, product, or style recognizable.
Veo3o1 separates those goals into First & Last Frames and Reference Images. Choosing the right mode before writing the prompt saves credits and avoids fighting the model with instructions that the selected inputs cannot support.
The short answer
- Choose First & Last Frames when the starting composition, ending composition, or transition between them is the main requirement.
- Choose Reference Images when a recurring subject or visual identity matters more than an exact first or last frame.
- Choose Text to Video when you do not need to preserve an uploaded visual.
First & Last Frames: control the endpoints
This mode accepts one or two images. With one image, you give the model a visual starting point. With two images, you define both ends of the shot and use the prompt to describe a plausible transition.
Good uses include:
- turning a product still into a reveal;
- moving from a wide scene to a close composition;
- creating a day-to-night transition;
- connecting two designed frames for a storyboard;
- animating a still while protecting its initial composition.
The two images should be visually compatible. If the subject, camera angle, lighting, and environment change completely, an eight-second transition may spend most of its motion trying to reconcile the conflict.
A strong transition prompt describes the path, not just the destination:
Begin on the first frame. The camera moves forward through the open doorway as evening light gradually replaces daylight. End on the second frame with the sign centered and the camera at rest.
See the First & Last Frames setup guide for the exact upload workflow.
Reference Images: control identity and visual language
Reference Images mode accepts one to three images. It is available with Lite and Fast, currently generates an 8-second result, and requires either 16:9 or 9:16.
It works well for:
- a recurring character;
- a product with distinctive shape or packaging;
- a consistent costume or color palette;
- a vehicle, room, or object that should remain recognizable;
- several shots that share a visual identity.
The uploaded images guide the subject, but they do not lock the opening and ending compositions. The model has more freedom to stage the shot around the referenced look.
For identity-focused work, use references that agree with each other. A portrait, a wider outfit view, and a side angle are more useful than three nearly identical images. Our character consistency guide covers reference selection and prompt structure.
Side-by-side decision table
| Question | First & Last Frames | Reference Images |
|---|---|---|
| Main goal | Control the shot endpoints | Preserve a subject or visual identity |
| Images | 1–2 | 1–3 |
| Best for | Transitions, storyboards, still animation | Characters, products, recurring visual worlds |
| Current reference-mode limits | Depends on selected controls | Lite/Fast, 8 seconds, 16:9 or 9:16 |
Frames guide the opening and ending; they do not guarantee a pixel-for-pixel match. Review both endpoints and the movement between them before using the result.
Three examples
Product reveal
You have a clean package image and a final hero composition. Use First & Last Frames. Describe the camera movement, lighting change, and how the package reaches its final position.
Recurring spokesperson
You need multiple clips featuring the same person in different environments. Use Reference Images. Keep the subject description stable while changing one action or location at a time.
Mood-only concept
You want “a rain-soaked futuristic market at night” and have no visual that must be preserved. Use Text to Video. A reference image would add constraints without solving a real requirement.
Avoid mode mismatch
Do not ask Reference Images mode to reproduce a precise final frame; it is not an endpoint lock. Do not use two unrelated frames and expect a natural transition without describing the path. Do not upload a reference merely because you have one—every input should support the shot’s primary goal.