Image to video with AI: how a photo becomes a moving clip

This article is an overview based on vendor information, documentation and publicly available reporting, not a scored test of our own. Prices and features change quickly; the vendor's own terms always take precedence. We only score what we have used ourselves for weeks: those tools are listed under AI tools compared.

How image-to-video works

In image-to-video, a still image serves as the first frame and the model invents what happens next: the camera moves closer, hair moves in the wind, a product turns on the table. The result is usually a clip of five to ten seconds.

Compared with pure text-to-video this has one big advantage: the look is fixed. Colours, product, person and framing come from the image, not from the model's imagination. For brands that need a consistent appearance, image-to-video is therefore the more reliable route. Several tools also accept an end frame, and the model calculates the motion between the two.

The source image decides

Most failed clips fail because of the image, not the prompt. The aspect ratio should already match the destination, 9:16 for Reels and Shorts, 16:9 for the website; cropping afterwards costs resolution. The subject should be clearly recognisable and have room to move: a person who reaches the edge of the frame can hardly turn without the model having to invent body parts.

Text, logos and fine patterns are the most likely to distort in motion. If you need them, add them in the edit instead. Hands, glasses and teeth are the classic weak spots; an image in which they are still and clearly visible succeeds more often.

The source image does not have to be a photo. Many people first create it with an image generator such as Midjourney or Adobe Firefly and then animate it; the free routes for that are listed in our article Creating AI images for free.

Keep the prompt to motion

Because the image already defines the content, the prompt only describes what moves and how the camera behaves. A good prompt is short: "Slow camera push towards the cup, steam rising, background stays still." Describing the image again tempts the model to reinterpret it.

One motion per clip is the safest rule. Every current model understands camera terms such as pan, zoom, dolly or drone shot; English tends to give slightly more reliable results than other languages. What should stay still can be stated explicitly: "The person stays where they are, only the camera moves."

And because every attempt costs credit: test in the faster, cheaper quality level that many services offer first, then generate the successful variant at full quality.

Which tool for which job

Kling is strong when characters need to look the same over longer clips. Seedance keeps products stable across several shots, which makes it interesting for advertising clips. Google Veo creates picture and sound in one pass, such as sound effects and speech to match the scene.

Luma Dream Machine is the easiest entry point and works with start and end frames. Runway brings the largest toolbox for everything that follows the clip, from extending it to replacing parts of the image, and Higgsfield relies on prebuilt camera moves that look more cinematic than a free description.

The differences only show on your own subject. Since almost every service has a small free or starter credit, it is worth trying the same image with the same prompt in two or three tools. What is possible without a subscription is covered in our article Creating AI videos for free, and individual portraits of the tools are in our video AI section.

Bringing old photos to life

One of the most popular uses is private: the grandparents' wedding photo that suddenly smiles and turns around. That can be moving, but it has two sides. The model invents expressions and movements that never happened; a memory becomes a staging.

Legally: anyone who animates living people and shows the result needs their consent. Nor are you free with the deceased; under the German Art Copyright Act, publishing a person's likeness requires the consent of their relatives for ten years after death. And in practice, family photos end up on servers in the United States or China when uploaded. If you only want to look at them privately, choose a vendor that does not use inputs for training, or switch that off in the settings.

Rights and labelling

The source image must belong to you or be licensed for this use. Some stock agency licences restrict processing with AI; a look at the terms saves trouble. As a rule you may only use the result commercially on a paid plan, almost never on the free tiers.

Since 2 August 2026 the AI Act has required that realistic, artificially generated or manipulated footage of people, places or events is disclosed as such. An animated product photo is unlikely to fall under this, a talking portrait of a real person certainly does. The video and voice tools we have tested and scored are in the AI video generators & voice AI category.

In short: the source image matters more than the prompt: matching aspect ratio, a clear subject with room to move, no text in the image. The prompt describes only motion and camera. Kling for characters, Seedance for products, Veo for sound, Luma to get started, Runway for post-production. Photos of people only with consent.