Veo 3 image to video prompts: frames, references, extend

A Veo 3 image to video prompt describes what happens after your picture, not the picture itself. Veo uses the image as the first frame, so the text should cover motion, camera and sound. With Veo 3.1 you can also set a last frame or give up to three reference images, and each mode wants a slightly different prompt.

This guide covers those three modes with Google’s own example prompts quoted word for word, plus the limits that decide which mode you can use in Google Flow and the Gemini API. For writing a text-only Veo prompt from scratch, start with our Veo 3 prompt structure guide.

Three ways to start a Veo video from an image

Google’s Gemini API Veo documentation lists all three for Veo 3.1: “Generate a video by specifying the first and last frames” and “Use up to three reference images to guide the content of your generated video.”

Image to video: write what moves

Google’s advice in the same documentation: “Veo uses the input image as the initial frame. Select an image closest to what you envision as the first scene of your video.” Its example uses a still made with Nano Banana as the starting frame. The image prompt was:

A hyperrealistic macro photo of tiny, miniature surfers riding ocean waves inside a rustic stone bathroom sink. A vintage brass faucet is running, creating the perpetual surf. Surreal, whimsical, bright natural lighting.

The video prompt that went with it:

A surreal, cinematic macro video. Tiny surfers ride perpetual, rolling waves inside a stone bathroom sink. A running vintage brass faucet generates the endless surf. The camera slowly pans across the whimsical, sunlit scene as the miniature figures expertly carve the turquoise water.

Compare the two. The video prompt repeats the subject in a few words so the model keeps it, then spends the rest on action (“ride perpetual, rolling waves”) and one camera move (“slowly pans”). A simple template built on that pattern:

[Style] video. [Subject from the image, in one line]. [What the subject does]. The camera [one movement]. [Sound: ambience, effects, or a line of dialogue in quotes].

The code sample on the same page goes even shorter and reuses one line for both steps: Panning wide shot of a calico kitten sleeping in the sunshine.

First and last frames: describe the transition

With a start and an end image, Veo fills in the motion between them. Google’s documented example is a ghost story:

A cinematic, haunting video. A ghostly woman with long white hair and a flowing dress swings gently on a rope swing beneath a massive, gnarled tree in a foggy, moonlit clearing. The fog thickens and swirls around her, and she slowly fades away, vanishing completely. The empty swing is left swaying rhythmically on its own in the eerie silence.

The prompt names what changes between the frames: the fog thickens, the figure fades, the swing keeps moving. In Google Flow the steps are the same idea. Flow’s help page says to choose Video, then Frames, drag images to “+ Add start frame” and “+ Add end frame”, and “describe the action or transition that should happen between the frames.”

Google’s prompt guide also shows that the end frame can come from a second image prompt. In its example, the first frame is A high quality photorealistic front image of a ginger cat driving a red convertible racing car on the French riviera coast. and the last frame is Show what happens when the car takes off from a cliff. The video prompt between them is marked “Optional”.

Reference images: say how each one is used

Reference images keep a subject consistent instead of fixing a frame. Google describes them as “up to three asset images of a single person, character, or product. Veo preserves the subject’s appearance in the output video.” Two of its examples:

Create a silly cartoon version of the fish wearing the costume, swimming and waving the wand around.

That prompt works with two references: an angler fish and a costume. It names both and says how they combine. Flow’s own example does the same with three ingredients, a woman, a lava lamp and a foggy street:

The woman, whose torso is the lava lamp, walks down the foggy street

Flow adds two practical tips: provide “subject or product references on a plain or segmented background,” and make sure “your prompt references the frames or ingredients you’ve provided.” A text prompt that contradicts the images works against them.

Extend a clip

Extension continues a Veo video you already made. Google’s API documentation says each extension adds 7 seconds, up to 20 times, and that “Extend finalizes the final second or 24 frames of your video and continues the action.” Its example extension prompt is a single line:

Extend this video with the paraglider slowly descending.

One catch from the same page: “voice is not able to be effectively extended if it’s not present in the last 1 second of video.”

Limits that change your prompt

From the Gemini API parameters table:

• Length: 4, 6 or 8 seconds, but 8 seconds only “when using extension, reference images or with 1080p and 4k resolutions”.
• Aspect ratio: 16:9 (default) or 9:16.
• Reference images: up to three in the API, on Veo 3.1 and Veo 3.1 Fast, not on Veo 3.1 Lite.
• People: for image-to-video, frames and reference images, personGeneration accepts only "allow_adult".
• Prompt length: 1,024 text input tokens per request.

In Flow, the models page lists first-frame and first-and-last-frame video on Veo 3.1 Lite, Fast and Quality, while ingredients (references) are supported on Lite and Fast for 8-second videos only, not on Quality. Gemini Omni Flash in Flow supports frames and ingredients at 4 to 10 seconds.

In the Gemini app, video generation now runs on Gemini Omni rather than Veo. Google’s Gemini Apps help says you can upload one video and up to five images, that the video “will match” the aspect ratio of an uploaded reference, and that you “must have rights to any images that you upload.” For plan credits in Flow, see our pages on Google AI Pro and Google AI Ultra.