Generating video

The body is where you describe the picture: subject, action, how the camera moves, what the light does. Type @ to reference the reference images already wired in (they're called "Image 1", "Image 2" in order, matching the badges on the thumbnails).

The button reads "16:9 · 720P · 5s · 1 clip", and the speaker at the end of it shows whether sound is on or muted. Open it and there are five groups:

When you change model, a tier the new model doesn't take (2K, say, going from H3 back to Seedance) is dialled back to the first tier it does take — nothing is sent out that the other side won't accept.

The number to the left of the send button is what this run will cost in credits, worked out live from duration × resolution × count, and it moves as you change the settings.

The rest of the buttons on the bar:

• Marker — circle the spot you mean on an image that has already rendered, so the model knows which element in the picture you are talking about (see Marking a local element below).
• Effects — opens the effect library: everything in there is a clip look published by someone else (or by you). Pick one and the chip folds away, the row below it turns into that effect's cover (hover it and it reads Replace; the × in its corner takes the effect off), and when you generate, its effect description is appended to your line while the model, aspect, resolution and duration are swapped over too. Picking one also drops an effect sample node to the left of this node (the sample clip if there is one, otherwise the cover) and wires it in as a reference — it really is reference material for this run, not just decoration. Publish effect on the video node's toolbar is how things get in there.
• Camera moves — opens the camera-move library: three tabs (Plaza / Favorites / My moves) plus search, with the 23 official moves (push in, pull out, pan, track, orbit, handheld…) coming with preview clips that play on hover and freeze back to the cover frame when you leave — what you see while choosing is what the move looks like. Clicking a card inserts a camera tag at the caret in your prompt (the tag carries a camera icon plus that move's name — select it and press Backspace / Delete to remove it): at render time that sentence lands where the tag sits, no longer appended at the end. A prompt takes at most three camera moves — click a fourth card and all you get is "up to 3 camera moves". Frequent picks go into Favorites; your own wordings live in My moves (a name, one line of description, and a cover — uploaded or picked from your history). The Camera moves button on the input bar is outlined whenever the prompt carries a tag, so one glance says whether this run has a move in it.
• ✨ Optimize prompt — the button in the right slot of the input bar, and its tooltip reads Optimize prompt. One click does not open a dialog: the button spins in place (you can keep typing the whole time), and when it finishes a blue dot sits on its top-right corner; click once more to bring the result up, under the title Optimized prompt. It rewrites what you wrote into a multi-shot prompt you can send straight off: a subject definition, then one line that refers to the subjects and restates the action, then one line per shot (shot size, what happens, camera movement, sound), and one closing line for the overall look; how many shots follows the duration picked on the input bar — 2-3 for 6 seconds and under, 3-4 for 7-10 seconds, 4-6 above that. The dialog offers Copy, Regenerate and Fill the input box: that last one replaces the whole input box with the result. Closing the dialog fills nothing — the result and the blue dot both stay, and opening it again shows the same one; change the text in the box and that result is stale, so clicking the button again regenerates from what is in the box now. Clicking it with an empty box asks you to write a prompt first. Your @ references and camera tags never get dropped — if the model misses one it retries once, and if it still misses one it reports that the optimization lost a reference instead of handing you a prompt that quietly lost it.
• Advanced — the button in the right slot of the input bar: a few switches, each a full-width pair of buttons, On / Off. How many rows you get depends on the model: four for the Seedance family (Web search / Check reference material / Smart reference / Watermark), two for Wan (Smart expansion / Smart reference), and Smart reference alone for everything else.

The Marker chip at the head of the input bar is the same thing as the one on the image input bar: a picture usually holds several things, and writing "the pillar in that picture" does not tell the model which one you mean — mark it on the picture and it knows. Clicking Marker puts you in Element selection: a bar floats at the top of the canvas reading Element selection and Click an image to select a local element, with Back to node and an × beside it, and the other nodes on the canvas all fade.

Click the spot you mean on an image node that has already rendered. There is a short wait between the click and the result, and the bar's second line reads Detecting… while it runs; once it comes back, a selection box and a name label are drawn on the source image, and clicking that label lets you pick another name and box from the candidates right around it. Clicking something that is not a rendered image (a draft, one still generating, a video node) says Click a generated image node, and a spot nothing can be recognized in says No element recognized here.

A successful mark does two things at once: a marker tag (the source image's thumbnail plus the element's name) is added to the end of the prompt on the input bar, and the image you clicked is wired in as a reference for this video at the same time — the same thing as drawing the line by hand, and it shows up where Using a reference image puts it.

When you send, that tag is not handed to the model as it stands: it opens out in place into a line saying which image and roughly where — the element's name plus its approximate area on that image. The same node expands the same way whether you send it from the canvas or from the command line / an Agent: there is one implementation of this and both render paths go through it.

The two sides are two ends of one operation: delete the tag in the prompt and the box on the source image goes with it; delete the box on the source image and the tag in the prompt goes too. One image can carry several marks, one tag per spot.

Wire an image node into the port on the left of the video node and that image is the reference for this generation — usually the first frame. You can also hit Reference on the input bar and then click the image you want on the canvas.

Images you've wired in line up as thumbnails above the input bar; double-click a thumbnail to jump the canvas to the node that image lives on.

The thumbnail has a mark in each corner: top left is its number among material of the same kind, matching the "Image N" in the prompt (in First & last frame mode the first and second carry First and Last here instead of a number); bottom left is what this piece of material is — a clip shows its duration, and an image taken only for its style carries a style mark; bottom right is the compliance check, spinning while it runs and a blue shield once it has passed. Hover the thumbnail and the top-left, bottom-left and bottom-right marks all fade out together, making room for the × at top right (drop this reference) and the @ at bottom right (insert its reference at the caret); move away and they come back.

There is one thing to watch for once a clip wired in counts as reference material: when it is shorter than the duration this run will render, a line above the input bar says plainly that the source is however many seconds and this run is however many seconds, and warns that it may be stretched automatically when generating. The button beside it dials to a tier that source can hold up: if the current model has a tier no longer than the source, only the duration comes down (the button reads Use a compatible duration); if there is no such tier at all, you get a model that can hold it (the button reads Use a compatible model). The line is a warning only and does not hold the submission back — whether to stretch is left to the model: render a 5-second clip as an 8-second output and the extra three seconds are made up by the model, and you only find out watching the finished cut. The line does not appear when the source duration cannot be probed — a missing number is no reason to raise it.

Characters on the input bar involves uploading or using someone else's likeness, so the first click brings up a usage pledge — media rights, the consequences of what you generate, and the consequences of breaking the rules. Tick to agree and it stops coming up. Agreeing takes you into the character library: pick a character and click Add to canvas, and its several reference sheets land on the canvas together, ready to be wired in as references for this run.