How to use reference-to-video AI (combine images into one clip)

Image-to-video animates one picture. Text-to-video invents a scene from nothing. Reference-to-video composes several source images into one. If you want a specific person wearing a specific outfit and holding a specific object, and no single photo has all three, R2V is built for exactly that: it treats each reference image as one ingredient and assembles them into a scene your prompt describes.

The trade-off is control at the cost of setup. You need clean, single-subject reference images and a prompt that correctly numbers and points to each one. Get the numbering wrong and the wrong element lands in the wrong place, a jacket on the wrong person, a face in the wrong body.

On the more capable engines (Seedance 2.0 is the clearest example), a "reference" isn't limited to a still image. You can feed in a video reference and recast its action or camera move onto a different subject, or hand it an existing clip and use reference-to-video as an editing tool: swap out a character for one of your reference images, or restyle the clip, while the rest of the footage stays untouched. That reframes R2V as a video editor as much as a video generator.

Audio references work the same way: on models that accept one, you can point an audio file at an image and use it to drive the animation, the classic case being a voice track that animates a still portrait's mouth and expression. It's a different entry point from writing motion in words: the sound itself supplies the timing.

Most reference-to-video models don't understand abstract feelings; they render body language. Writing "she feels nervous" gives the model nothing concrete to animate. Writing "she is glancing away and biting her lower lip" gives it an actual physical action to render.

This holds for every emotion in the prompt: translate the feeling into a micro-movement or body cue before you write it down. It's the difference between a shot that reads as flat and one that reads as alive.

Wrong element in the wrong place: your reference number doesn't match the actual upload order. Recheck the mapping.

Identity or outfit drifts partway through the clip: add a second angle of that subject as another reference and name explicitly what to preserve ("preserve the jacket exactly from reference 2").

Style doesn't match the references: spell out the style, palette, and lighting explicitly in the prompt, and repeat the style block on every scene if it's a multi-shot sequence.

Motion feels weak or the shot looks static: add an explicit action verb and a named camera move, the same fix as any video prompt.

For anything with more than one beat, don't cram it into a single sentence. Open with a reference setup line and a global style block, then break the rest into shots with timecodes: "SHOT 1 (0:00-0:05): [camera]. [action]. SHOT 2 (0:06-0:10): [camera]. [action]." Keep each shot to one clear action, and repeat the style block if the scene changes so the look doesn't drift across cuts.

Some of the more advanced engines take this further and let you script a full scene as structured, near-JSON shot data instead of plain prose, camera moves and cuts defined field by field rather than described in a sentence. It's the same discipline as the timecode format above, just in a stricter shape, and it's worth checking whether your specific tool offers it before you hand-write a long multi-shot prompt.

If one photo already shows everything you want animated, plain image-to-video is simpler and has one less thing to get wrong. Reach for reference-to-video specifically when the shot you're after is assembled from parts, a person from one photo, an item from another, and no single image contains the whole scene.

The techniques also chain: use reference-to-video to generate a combined starting frame or short clip, then treat its output as the source image for further image-to-video animation, or as one more reference in a later R2V pass.