Start and end frames, and reference clips

On Video, the bar has a tab for each way of making a clip: Text / Image to Video takes frames, beside the prompt, and Reference to Video takes references, in a box above it. Each tab's model menu lists the models that make from what it takes, and the slots follow the model you pick: there's a slot for every kind of input it reads, and none for the kinds it doesn't. In the model menu, each row's badges say what it takes, such as "Start & end frame", "9 image refs" or "3 audio refs", and the Takes filters at the top narrow the list to the models that take what you have.

What you put in one tab stays there while you're on another, and only the tab that's open goes with Generate.

Start and end frames

These are on Text / Image to Video, the Start and End tiles beside the prompt.

• Start frame: the clip begins on your picture. Describe what moves: the action, the camera, the light.
• Frames: a start frame and an end frame, and the clip travels from one to the other. Describe how it gets there. An end frame needs a start frame beside it.

The Seedance models, Kling 3.0, Kling 4.0 on Keyframes, MiniMax H3, PixVerse v6, Wan 3.0, Wan 3.0 Prime, Gemini Omni 1.1 and Grok Imagine 1.5 take both. Kling 3.0 Turbo, Kling 4.0 Flash on Keyframes, Grok Imagine Video, PixVerse v6 Max and HappyHorse 1.1 take a start frame. Vidu Q4 takes a start frame or image refs, and needs one of them. A clip made from a start frame on Kling 3.0 Turbo, Kling 4.0 Flash, PixVerse v6, PixVerse v6 Max, Grok Imagine Video, Grok Imagine 1.5, HappyHorse 1.1, Vidu Q4 or MiniMax H3 takes the picture's shape, and on Seedance 2.5 and Kling 4.0 the Shape menu offers one choice, adaptive or auto, which fits the clip to your picture.

To animate a picture you've already made, open it in your Library and press Animate: Video opens with it in the start frame.

Kling 4.0: Keyframes or Omni

Kling 4.0 runs one of two ways, and the tab picks which:

• Keyframes, on Text / Image to Video: a start frame, an end frame, and up to 8 frames between them, in order. They sit between Start and End, tagged @Keyframe1, @Keyframe2 and on. A clip takes one more frame in all than its whole seconds, up to 10, its start and end among them: a 5-second clip with both takes 4 between. Frames between need a start frame, and the bar says so if one is missing.
• Omni, on Reference to Video: up to 10 image refs and 5 video refs, and no frames.

Each keeps its own: the frames between stay on the first tab, and the image refs on the second.

Kling 4.0 Flash has the same two tabs, at 720p and up to 20 seconds. Its Keyframes takes words or a start frame, with no end frame and no frames between; its Omni takes up to 10 image refs and 5 video refs, each video 2 to 20 seconds and 20 seconds in all.

Grok Imagine 1.5: Frames or References

Grok Imagine 1.5 runs one of two ways too, picked by the tab:

• Frames, on Text / Image to Video: a start frame, an end frame, and up to 4 frames between them, in order. Each goes at a whole second inside the clip, spread evenly, so a clip takes one fewer than its seconds: a 5-second clip takes 4, a 3-second clip 2. With a start frame and no end frame, the bar has a Loop button: the clip ends on its start frame, so it plays round. It doesn't change the price.
• References, on Reference to Video: up to 7 image refs, or one at 1080p.

On a board it works the same way: the clip card has the Mode menu, the stills on its Image refs port are frames between on Frames, and Loop sits beside Sound while only a start image is wired.

References

These are on Reference to Video, in the box above the prompt. Its head counts what's in of what the model takes, such as "2 / 9 pictures", and the ⓘ beside the counts says what the model takes and how long its clips and sounds may be.

• Image refs: people, products, places or a look for the clip to keep.
• Video refs: clips for the new one to follow. A video ref the prompt doesn't name is followed for its motion and camera; the people in it don't come with it.
• Audio refs: a voice or an ambience for the clip.

Each one gets a tag in the order you add it: @Image1, @Image2, @Video1, @Audio1 and on. Type @ in the prompt to pick one from a list, and say what each is for, such as "@Image1 walks through the market in @Image2, the camera moving as in @Video1". Seedance, Wan 3.0, Kling 4.0 and Kling 4.0 Flash read the tags; for other models they're turned into words, so you can use the same tags whichever model you pick.

Once a clip or a sound is in, the line under the box says how many of them the model takes and how long they may be, such as "Up to 3 · each 2–15 s · 15 s in all". A few rules hold across models:

• Audio refs go beside an image or video ref. Wan 3.0 and Wan 3.0 Prime also take a sound on its own.
• Kling O3 (reference) follows a video ref only with an image ref beside it, and makes no sound with a video in.
• On Wan 3.0 and Wan 3.0 Prime, the video refs and the clip you're making add up to 30 seconds at most.
• Kling O3 (reference), under From references only in the menu, needs at least one ref. Vidu Q4 takes audio refs only beside an image ref, as MP3 only, each 3 to 12 seconds.
• Kling 4.0 and Kling 4.0 Flash take no audio refs for now. They make their own sound with every clip.
• Grok Imagine 1.5 takes up to 7 image refs at 480p and 720p, and one at 1080p. With more at 1080p, the run is turned down before anything is charged, and says to pick a lower resolution or take some out.
• Kling 4.0 and Kling 4.0 Flash take image refs from 2:5 to 5:2 in shape. On Video, Generate stays off for a taller or wider one and says which. On a board, the clip card says which picture it is and Run stays off, and so does a start image that goes in as the opening scene; nothing is charged. Pictures in other formats, such as WebP, are turned into JPEGs or PNGs for it, since it reads only those.

If something doesn't fit, Generate stays off and the bar says what to change, such as how many to take out.

The opening scene

A model takes frames or references in one run, not both. On Reference to Video, the box's Opening tile takes a picture for the clip to open on, on models that take a start frame: it goes in as one more image ref, the last one, as the clip's opening scene. Its tile shows the tag it goes in as, such as @Image3, and it counts toward the model's image refs. It's the same picture as the Start frame on Text / Image to Video, so a start frame you've added there is the opening scene here. An end frame stays on Text / Image to Video: a clip made from references doesn't keep one. On Wan 3.0 and Wan 3.0 Prime, the opening scene is kept aside while a page or a document is in.

Adding files

Each tile opens a chooser with your Uploads and what you've Made; a picture tile also has Saved, for your Cast and products. From your work, at the top right of the references box, opens the same chooser for the kind you pick. You can drop a new file on the chooser, on the tile, or anywhere on the bar, and it lands in the slot that fits. A video or a sound dropped on Text / Image to Video goes to Reference to Video, where the model takes it, and that tab opens. Pictures can be PNG, JPEG, WebP or GIF, videos MP4 or MOV up to 200 MB, and sounds MP3 or WAV up to 15 MB. A voice note from your phone (M4A, AAC, OGG or Opus) goes in as an MP3. Vidu Q4 takes MP3 only, and the bar says so before you press Generate. You can play a sound from its tile in the chooser before you pick it.

A sound longer than the model takes is marked on its tile. Press the scissors on it to trim it: drag the part to keep, play it on a loop, cut the silence at the ends, or trim it to the length the model takes. The trimmed part takes the track's place in the slot, and the original stays as it was.

If you switch to a model that doesn't take something you've added, it stays in its slot, kept aside and unsent, until you pick one that does.

Whose voice it is

When you upload a recording as an audio ref, the bar asks whose voice is in it: a real person, named, with their consent confirmed, or no one speaking or singing. Generate stays off until you answer. Voiceovers and songs you made on Vessero aren't asked about.

What references cost

Seedance, Wan 3.0, Wan 3.0 Prime and MiniMax H3 charge for the seconds of reference video you send, on top of the clip. Kling 4.0 and Kling 4.0 Flash charge a lower rate a second with a video ref in, on the clip's seconds and the video's, and counts a video shorter than the clip as the clip's length. Image and audio refs add nothing. The price on Generate includes it all; hover over Generate to see the seconds counted. You're never charged more than that price.

See them again on the finished clip

Open a finished clip in My videos, or in your Library, and its frames and references are listed under its prompt as References, each by the name your prompt gave it. Press one to see or play it in the clip's place, with ← and → to step through the rest; Back to the clip, or Esc, brings the clip back. A web page or a document a clip was made from opens in a tab of its own.

On a Creative Board, a clip card takes frames and references through wires instead: see Next shot, frames and reference videos on a board.