Image-to-video

In short

Image-to-video is AI that turns a still image into a moving video clip, usually guided by a text description of the action or camera movement. It adds plausible frames while trying to preserve the pictured subject and composition.

Also called: Image-to-video generation, I2V
Topic: Images, video, voice and music
Level: Beginner
First described: Zhang and colleagues introduced I2VGen-XL, a named high-quality image-to-video synthesis system, in November 2023.

An image-to-video model takes a still image as a starting frame or visual reference, then generates a sequence of frames that depicts movement over time. The accompanying prompt normally describes the motion, camera work or change, rather than repeating what is already visible¹.

Many current systems use diffusion models. They learn patterns from video and image data, then produce frames while conditioning the result on the input image and text. I2VGen-XL, for example, uses a base stage to preserve content and coherent semantics, followed by a refinement stage for detail and resolution².

The result is not a recording of what happened after the photograph was taken. It is a new sequence, so hands, faces, text and physical movement can change or become inconsistent. A still image can guide composition, subject, lighting and style, but it does not guarantee that every generated frame will remain faithful¹.

You meet the technique in creative video products. Runway documents image-to-video in Gen-4.5 and asks users to upload an image and describe the desired motion¹. It is used for animating a product still, making a character move, creating a camera push-in, or turning a family photograph into a short visual effect.

Example

Upload a photo of a paper boat on a puddle and prompt, "the camera slowly follows as the boat drifts past floating leaves". The system creates the intervening motion; it does not recover a real recording of that boat.

Why it matters

It lets people prototype moving shots from artwork, photographs or storyboards without filming every frame. The trade-off is that the generated motion is an interpretation, not evidence of a real event.

Questions people ask

What is the difference between image-to-video and text-to-video?

Image-to-video starts with a still image and uses text mainly to specify motion or camera behaviour. Text-to-video starts from a written description and must create the scene as well as its movement¹.

Can image-to-video animate any photograph?

It can attempt to animate many photographs, but the output may distort faces, hands, lettering or object geometry. The input image provides guidance, not a guarantee of frame-by-frame consistency¹.

Is image-to-video the same as making a time-lapse?

No. A time-lapse joins real photographs captured at different times. Image-to-video generates new frames between or after the visual information in a still image, so its motion may be invented.

Related terms

Tools that use it

Sources

Back to the glossary