Skip to content
Docs
Docs

Making videos

Making videos with the Assistant from words, an image or several images, or a video's motion, playing them on the Canvas, and trimming, editing, upscaling and cutting them out in the video editor.

Markdown

Choose Video from the mode button at the start of the prompt (see Talking with the Assistant) and describe it, such as “waves at dusk, slow and calm”. The mode decides what is made, not the words: in Image, “a film still of waves” is an image. Right-click an image, choose Use as reference, and describe how it moves (“the light sweeps across the sea”): the video starts from that image and appears beside it, in its shape; the Composer says “Animating Image 1”. Choose several images and the video is guided by them. Choose an image and a video, and the image’s character moves as the video moves (see A video’s motion). In an image’s editor, Video animates that image, guided by any references you’ve chosen: they are Image 1 onwards, in the order you added them, and the image itself is never one of them. In Ask, “make this into a video” has the Assistant make one.

Like an image, the video’s place appears at once and glows while it is made (usually a minute or two), then shows its first frame. Until then, Stop in the Composer cancels it at no cost (see Stop). Press the play button at its bottom left to play it on the Canvas, muted and looping, and press it again to pause: it carries on from where it paused. While you hover the video, a small scrub bar beside the button shows how far it has played: press or drag along it to jump to any moment, playing or paused (with the bar focused, the arrow keys move five seconds). On a video shown too small for it, there is only the play button. It keeps playing while you pan, zoom and select. On the Canvas, a video plays at its full quality, up to 4K, on every plan; exports and the models use the original, as it was added. When it leaves your view, or you switch to another tab, it pauses where it is, and waits for you to play it again.

Select a video and its properties panel at the Canvas’s top right has Play or Pause; with several videos selected, Restart sits beside it, and each says how many videos it acts on. Restart, Loop and Mute are also in the panel’s ⋯ (More), and all of it is in the video’s right-click menu: Play, Pause, Restart, Loop and Mute. Turn Loop off to play it once: it stops at its end, and plays again from the start. Loop is saved with the video, so everyone sees it play the same way. Playing, pausing and sound are yours alone: pressing Play never plays anything for anyone else. Select several videos and Play starts them all together from the start, each looping or playing once by its own setting; Pause and Restart act on all of them, and other videos carry on as they were. In a selection of images and videos, these act on the videos only. It is an ordinary node: move, resize, flip, copy or delete it. If it can’t be made, it turns red like an image does, with Retry. Once it is ready, right-click it and choose Download original to save its MP4 exactly as it was made. People viewing a shared link can watch it, but have no export.

Choose a video’s shape, resolution, length and sound in Settings, the sliders button beside the cost meter (see Settings). Its shape starts on Auto: a video made from an image takes that image’s shape (an image you animate, or the first one you chose), and one from words alone is 16:9; choose a shape and it stays, whatever images you choose. The Assistant uses MiniMax H3 Turbo unless you choose another in Settings: the fastest and cheapest, it always makes sound, and offers 480p, 768p and 1080p. The others:

  • MiniMax H3 Max: MiniMax’s quality tier, at twice Turbo’s price.
  • Seedance 2.0 Fast: fast, with sound you can switch off.
  • Wan 3: up to 30 seconds long, with sound you can switch off.
  • Kling 3 and Kling 3 Pro (1080p): cinematic, with sound you can switch off, in 16:9, 9:16 or 1:1.
  • Seedance 2.0: higher quality than Seedance 2.0 Fast.
  • MiniMax H3 Lite and MiniMax H3 Highspeed, labelled Experimental: offered to try out, not yet tested by us.

A video made from a video’s motion uses Kling 3 Motion Control, or Kling 3 Motion Control Pro if you choose it (see A video’s motion).

Where a model can, Settings also offer Enhance prompt (it writes your prompt out more fully before making the video; on unless you switch it off). A video made from one image keeps that image’s shape. A video is priced by its length and resolution, and the cost meter beside the send button follows what you set (see Usage and prices).

On Free, AI video is made at up to 720p, and MiniMax’s models also at their own 768p. Larger resolutions show in Settings, greyed out, with the plan that includes them. A model that only makes larger video (Kling 3 Pro, at 1080p) needs a paid plan.

Downloading a video

Right-click a video and choose Download original to save the file exactly as it was added or made (“Clip (original).mov”), or Download video… for its other sizes. Or select it and open Export at the foot of its properties: choose Original (where it starts) or a resolution (add rows with + for several) and choose Export, and each is made and downloaded the same way; the resolutions you chose are kept with the video (see Properties). In the library, choose Download on one video. Download video offers:

  • Original: the file as it was added, at its own quality. Always there, on every plan, even if the other sizes can’t be found (then the dialog says so, with Try again).
  • Full resolution: for a video that was converted (a MOV, WebM or MKV, or an HDR video), the H.264 MP4 Otherspace plays and edits; for one that is trimmed or has its audio removed, an MP4 of what it plays, made from the original. On Free it is there only for videos of 720p or smaller.
  • 4K, 2K, 1080p, 720p and 480p: sizes smaller than the video, each an H.264 MP4 made from the original when you ask, keeping its shape, the right way up, with its sound in step (or none, once its audio is removed). A size your plan doesn’t include shows the plan that does: on Free, exports are up to 720p.

An export carries on while you work, even if you close the dialog or the page. It shows in the panel at the bottom right, where you can cancel it, and downloads when it’s ready, with Download there in case your browser didn’t take it. Asking for the same size again gives the same file. A finished export is kept for 7 days after you last download it. Exports don’t count towards your storage. On Free you can export 5 minutes of video a day; downloading originals is never counted.

First and last frames

With images chosen, and a model that starts or ends on an image, a panel above the Composer shows what the model is sent: First frame and, where the model can end on an image, Last frame, then References, and what is Not sent. Drag an image from one to another, or open its menu and choose First frame, Last frame, Reference or Don’t use. Esc or the arrow folds the panel away; Roles, beside the images, opens it again, and you never need it open to send. Under each image, the Composer says what it is: First, Last, or Not sent. An image selected on the Canvas is the first frame to begin with (see Creating an image on the Canvas), shown as This image, never numbered.

On Auto, one image is the first frame. Two images are the first and last frames for Kling 3 (which takes no other images), and guide the video for the other models. Three or more always guide it. Choose a frame and the rest follow: an unchosen first frame is Image 1, and an unchosen last frame is what Auto would end on. To have one image guide a video rather than start it, make it a Reference.

What each model takes:

Model First frame Last frame Other images beside the frames
Seedance 2.0 and 2.0 Fast Yes With a first frame Not sent
Wan 3 Yes With a first frame Not sent
MiniMax H3 Turbo Yes Yes, even alone Not sent
MiniMax H3 Max Yes Yes, even alone Sent as references
MiniMax H3 Lite and H3 Highspeed (Experimental) Yes With a first frame Not sent
Kling 3 and Kling 3 Pro Yes With a first frame Not sent (Kling takes no references)

Where a model can’t use the other images beside its frames, you don’t need to take them out: they stay attached, shown as Not sent, and the Composer says why, such as “Wan 3 uses frames or references, not both. Image 2 isn’t sent.” Once the video is sent, they’re let go with the rest, so they never go silently with your next prompt. A choice the model can’t make, such as a last frame with no first on Seedance, says why before anything is sent. A video’s Metadata says which image was its first frame and which its last.

A video’s motion

A video can guide a new one: its motion (body, gestures, expression, a push-in of the camera) moves the character in an image. A dance clip makes the person in a portrait dance, in the portrait’s own setting.

Right-click an image and a video on the Canvas (or choose them in your library) and choose Use as reference, or select the image and choose Use as reference on the video. In Video, the panel above the Composer shows the image as the Character and the video as its Motion; the Composer says “Moving Image 1 as the video moves”. Describe the character and the scene: the moves come from the video. The new video appears beside the image, in the image’s shape, as long as the video.

  • Kling 3 Motion Control is used unless you choose another in Settings: about 720p, 30 frames a second.
  • Kling 3 Motion Control Pro is sharper, at about 1080p, and costs more. On Free, it shows the plan that makes 1080p.
  • The video’s own sound is kept. Switch Keep the video’s sound off in Settings for a silent video. Its shape and length aren’t chosen: they are the image’s and the video’s.
  • The video you give it can be up to 10 seconds long, and at least 3. A longer one is refused before anything is charged: trim it first (see Trimming), and only the part kept is followed.
  • It takes one image and one video. Another image or video stays attached, shown as Not sent, and the Composer says why. Other video models never offer the Motion role: a video attached to them isn’t sent.
  • It is priced by the video’s length in whole seconds, as the provider charges it: a 3.2-second video is priced as 4 seconds. The cost meter shows it before you send.
  • It takes about 2½ to 4 minutes. If the model refuses it (a likeness it won’t move, say), it costs nothing, and the video’s place says why.

In Ask, attach the image and the video and say what you want (“make this character do the moves in that video”): the Assistant makes it the same way, at the same price.

Videos are muted until you untick Mute. Not available yet: pasting videos, and videos in exported images.

Editing a video

Double-click a video on a Canvas, or right-click it and choose Edit, to open it in the video editor. The video grows from where it was into the middle of the screen, and the address becomes the video’s own (/videos/…); your Canvas stays just behind it. From your library, Open on a video opens it the same way, on its own. It works like the image editor: the video in the middle, its tools down the left, and beneath it the timeline.

Select a video and its properties panel at the top right also offers Trim, Upscale and Remove background: each opens the editor with that tool already chosen.

  • Play and pause with the play button above the timeline, by clicking the video, or with Space. Sound plays here unless you press the speaker to mute it.
  • Move to any moment: drag along the timeline, or click it. With the timeline focused, the arrow keys step a tenth of a second (a second with Shift), and Home and End go to its start and end. Once the video’s preview frames are ready (shortly after it is added), the timeline shows them along its length.
  • Zoom and move around as in the image editor: scroll to move, hold Ctrl (⌘ on a Mac) and scroll to zoom about the pointer, or drag the video. The zoom control in the bottom-right corner, the Canvas’s, zooms in and out with − and +, takes a typed level, and its menu fits it again. Ctrl with + and − zooms, with 0 fits, and with 1 shows it at its own size.

Trimming

Choose Trim in the toolbar (or press T), and the timeline shows the whole video, its ends now handles. Drag one in and let go: the trim is kept straight away. You can also press I and O to trim the start or end to the playhead. Put Trim down and the timeline is just the part you kept, timed from its start. Pick Trim up again and the whole video is back, with what is left out dimmed, so you can let a trim out again: drag both handles back to the ends and nothing is trimmed.

Trimming is free and changes nothing for good: the whole video is always kept, and a trim is only where it starts and ends. A trim keeps at least a second, and can be any length, of a video of any size. The Canvas shows it at once: it plays only the part kept, loops or stops at its end, and shows its length. No file is made when you trim. One is made, from the original, only when one is needed: when you download or export the trimmed video (see Downloading a video), or give it to a model, which waits for it a moment. Download original still downloads the whole file as it was added. Trimming doesn’t make a version (see Versions), and Undo and Redo go back and forth between trims at once.

Removing audio

Choose Remove audio, beside the speaker above the timeline, to take the video’s sound away. Like a trim, it is free and changes nothing for good: the video plays silent in the editor and on the Canvas (where it has no Mute), it downloads and exports with no sound at all, and a model given it gets no sound. A file without sound is made from the original only when one is needed, as a trimmed video’s is. Download original still downloads the file as it was added, with its sound. Choose Restore audio to bring the sound back, or use Undo. It doesn’t make a version, but a version you create keeps it.

Upscaling

Choose Upscale in the toolbar to make the video sharper and larger: 2× or 4× its size, each showing the size it makes and its cost meter before you choose. Upscaling uses Topaz Video Upscale.

While it is made (a minute or two), Upscaling… shows on the video itself. Then the result plays in the video’s place:

  • Accept makes it what the video is, wherever the video is. It doesn’t make a version.
  • Discard leaves the video as it was.

A step that would make it larger than 4K isn’t offered, and neither is upscaling more than a minute of video. A video that is trimmed and has its background removed can’t be upscaled yet: undo the trim first, or upscale it before removing the background. Like every model, it is charged as it is made, whether you accept it or not (see Usage and prices).

Editing

Choose Edit in the toolbar and describe the change (“change the horse to white”, “make it snow”). What you describe is changed, and the rest of the shot is kept as it was: its framing, motion and timing, and its length. It uses FLUX 3 Video Edit Fast, which makes video up to 720p, at 24 frames a second, and works on up to 14 seconds of video (trim a longer one first). It keeps the video’s sound, and takes about a minute for a 5-second clip. Its cost, by the length of the part kept in whole seconds, shows on the meter before you choose Edit video.

A video above 720p offers two buttons, so it never loses resolution unless you choose to:

  • Edit at 1080p (or the video’s own resolution) uses Kling O1 Edit, which keeps the video’s resolution. It works on 3 to 10 seconds of video, up to 2160 pixels on each side, and costs more.
  • Edit at 720p uses FLUX 3 Video Edit Fast, as above.

Each shows its own cost. On a plan whose AI video is up to 720p (Free), Edit at 1080p names the plan that makes it, and Edit at 720p comes first. Asked through the Assistant, an edit of a video above 720p keeps its resolution where your plan makes it, unless you ask for 720p, and says which before you agree.

While it is made (about 5 minutes), Editing… shows on the video itself. Then the result plays in the video’s place: Accept keeps it, Discard leaves the video as it was. An edited video is a new picture, so a removed background doesn’t carry over: remove it again if you need it. Like every model, it is charged as it is made, whether you accept it or not.

Removing the background

Choose Remove background in the toolbar to cut the video’s subject out. Its cost, by how many frames the part kept has, shows on the meter before you choose. It uses VEED Background Removal.

While it is made (usually under a minute), Removing the background… shows on the video itself. Then the cut-out plays in the video’s place, over a checkerboard where the background was. Accept keeps it, without making a version; Discard leaves the video as it was. To bring the background back, go back to an earlier version (see Versions).

On the Canvas the video shows cut out too, over whatever is behind it. A trimmed cut-out video isn’t shown on the Canvas yet: the editor says so, and the Canvas goes on showing the video as it was. More than a minute of video can’t have its background removed. Like every model, it is charged as it is made, whether you accept it or not.

Tools on a trimmed video

Upscale, Edit and Remove background work only on the part the trim keeps. Only that part is sent to the model, so you pay only for it: its cost, shown before you choose, is for its length (or, for Remove background, its frames). The whole video is still kept. A file of the part kept is made from the original first, so the model may wait a moment before it starts.

The result is that part, made again. Once you accept it, it is what the video plays, from the start of the part you kept to its end, and you can trim it further. The parts you left out aren’t in the result. To get them back, use Undo, or go back to an earlier version (see Versions): the video is as it was, trimmed as before, and you can trim it differently and run the tool again.

How the video was made

Once you have accepted a change, the steps that made the video are listed beneath its timeline, in order: its Original, then each change you kept, such as Upscaled 2× or Background removed. Open a step to see its settings, what it was made from, what it made (a removed background also makes a matte, the outline it is cut out by) and what it cost. While a step is open, the video shows as it was after that step; choose Back to current to return. Trims aren’t steps: they are where the video starts and ends.

Undo and redo

Press Ctrl+Z (⌘Z on a Mac) to undo your last change in the editor: a trim, removing or restoring the audio, or a result you accepted. Ctrl+Shift+Z (or Ctrl+Y) does it again. Undo never makes or removes a version, and going back to a version starts undo afresh.

Versions

A version is a moment you choose to keep: nothing makes one by itself. Versions, at the top, lists the video’s versions, newest first, starting with its Original, and ticks the one the video is while you haven’t changed it since. Choose Create version to keep the video as it is now.

Choose a version to go back to it. The version itself never changes, and anything you did since your last version is kept as Current at the top of the list, so you can go back to that too. Editing on from a version you went back to leaves your other versions as they were.

The video’s own conversation

The Assistant stays beside the video editor with the video’s own conversation, as an image’s does: ask about the video in Ask. Ask it to make the video sharper or larger (“upscale this”), to change it (“make it snow”), or to cut out its subject (“remove the background”), and it offers the editor’s Upscale, Edit or Remove background with its cost: nothing is made until you agree. Making something else from a video in Image or Video isn’t available yet.

Sharing

Share, at the top right beside the close button, shares the video by link, with the same dialog as Share… in the library: Restricted or Anyone with the link, Allow download, Copy link and Reset link. It shows for anyone who can share the video; someone who doesn’t manage its Space can copy its link but not change it. Someone opening the link sees the video in this same editor, view only (see Sharing).

Press Escape (with Trim put down), or the close button at the top right, to fly the video back to where it was.