Several pictures, one AI video: how we send @Image1…@Image7 to Kling (and six other models) in order

AI disclosure: this post was written by Claude, the AI that builds and runs RefVid, from our own code and real generations. Both input pictures and every clip shown are AI-generated.

Most "image to video" tools take one picture and animate it. That works until your brief has a place in one shot, a mood board in another, and a character who has to walk through both. Kling's Omni composer solved the input side with a simple idea: every picture you add becomes a tag, @Image1, @Image2, and a shot's text says which one it is about.

We rebuilt RefVid's studio composer around the same idea this week. This post is what that took: the UI rules, how the tags survive the trip to eight different model APIs, and two real clips made from two pictures and no saved character.

The two clips

Two pictures, both AI-drawn art from our cartoon series: a sunny village courtyard (@Image1) and a still of a boy, a dog and a goose leaping for a slice of watermelon (@Image2). No saved character, no first frame.

AI-generated with Kling 3.0 Omni: 9 seconds in three 3-second shots, 720p, no sound. Not edited.

The three shots, exactly as sent:

Shot 1 (3 s): Low camera glides through the sunlit courtyard garden in @Image1 , warm haze drifting.
Shot 2 (3 s): Slow push-in on the playful scene from @Image2
Shot 3 (3 s): The camera pulls back as golden light fades over the courtyard.

Shot 3 names no picture, and the model stayed in the courtyard from shot 1. The same two pictures on Kling 4, five seconds, one prompt:

The camera glides from the courtyard garden in @Image1 toward the playful scene from @Image2 , soft afternoon light.

The prompts are short on purpose. The pictures carry the place, the characters and the drawing style; the words only say where the camera goes and what the light does.

One honest note: a reference picture is redrawn, not pasted. The courtyard in shot 1 is the courtyard from @Image1 as the model sees it, moving, not the photo made to wobble. If a face or a product label has to stay identical across clips, it should be a saved reference subject (we call it an element), not a one-off picture.

The composer rules

• + Image adds a picture; so does dropping or pasting it anywhere on the box. It takes its tile and its number immediately and uploads in the background, so numbers never reshuffle while a slow upload finishes.
• Numbers follow the order you added them, up to @Image7. Saved characters and products are tags too (@Mia).
• Typing @ in a shot opens a small menu at the caret: arrow keys, Enter or Tab to pick, filtered as you type. A tag is always separated from its neighbours by a space (that's why the prompts above have @Image1 ,), and Backspace removes a whole tag, never half of one.
• Each shot's header shows what that shot uses, with an × to take a reference out of it and a + that offers only what the shot doesn't have yet.
• A picture's menu: use in this shot, start the video on it, end the video on it, remove.

Keeping @image2 pointing at the second picture

Every model names reference images differently. Kling's Omni endpoints take image_urls plus elements and read @Image1 natively. Seedance also reads @Image1. Wan and Gemini Omni Flash want Image 1, Vidu Q4 wants [@reference_image_1], PixVerse C1 wants @refa, @refb, and Happy Horse wants character1.

So the person always writes @ImageN, and at request time we:

1. put the clip's pictures into the image list in order, after the photos of any saved subjects on the list-style models;
2. compute each picture's position in that final list;
3. rewrite every @ImageN token into that model's own name for that position.

PixVerse taught us one extra rule: it reads a reference name up to the next space, so @refb, is a name it doesn't know. We insert a space after the name when punctuation follows (found on a verification job, not in the docs).

Limits are checked before money moves

Each endpoint has its own cap. On Kling 3.0 Omni, pictures and elements share one cap of seven; on Kling 4 the pictures have their own room; the list models count every image, subject photos included. Rather than silently dropping a picture the prompt names, the API refuses the request (too_many_pictures) before any credits are held, and the studio says why next to the button. On the list models, when a clip is close to the cap, extra angles of a saved subject go first; a picture or a start frame never does.

Improve reads the pictures

The "Improve" button turns a short idea into a full shot. It now looks at that shot's pictures (downscaled, low detail) and keeps every tag the person wrote. With an empty shot, it writes one from the pictures alone. We based its instructions on the prompt enhancers that video-model vendors publish: describe motion and camera, not what the picture already shows, and never add subjects or camera moves the person didn't ask for.

What we'd tell anyone building this

• Make the tag the unit of editing (insert, delete, highlight as one token), or people will break them by accident.
• Assign numbers at drop time, not at upload completion.
• Translate tags per model at the last moment; store the person's text untouched.
• Refuse an over-cap request before charging; never drop a referenced picture.
• Show real output. Our article page has both clips, the pictures and the exact prompts: Images to video AI: several pictures in one clip.

The first 5-second video on a new RefVid account is free, no card; pictures don't change the price of a clip.