The Master Guide / 07 Video with Seedance · Setting up a shot
Attaching references
You attach the smallest set of anchors that locks what must hold, each given one job.
You attach the smallest set of anchors that locks what must hold, each given one job.
Attach the identity set [proven for two images and the short anchor line]: a face that reads gets its sheet plus the identity close-up; a figure seen small, from behind or out of focus, the sheet only. Two identity images is the ceiling: never a third image of the same person, because near-duplicates average the face.
The smiling close-up. Option: use it when the script has a smile or laugh in a close or medium shot [single strong source]. That character gets a smiling close-up (Photoreal anchors (Nano Banana Pro) → Close-ups and casting stills). In close shots where the character speaks, sings or laughs, it may take the neutral close-up's place [best available, not yet validated].
Drawn, anime and 3D characters [best available, not yet validated]:
the style anchor, once per generation;
per character, the three-quarter design view in the sheet's role, and the turnaround's portrait (or the expression map, for broad acting) in the close-up's role;
the full turnaround only for turns, jumps and fights;
never a photographic face.
How they are made: see Stylised films (Seedream 5.0 Pro) → Turnaround, expression map, state sheets.
How many [proven]: usually 2–6, aim 4–6, about 8 is the warning line: one plate per location, one reference per hero product or prop in frame, nothing absent from the generation. Over budget, cut from the bottom of this order: identity close-ups, sheets, product, plate, hero prop, label close-up, coverage plate, small figures. If one keeps hijacking the shot, remove it.
Role lines: tags run @image1 to @imageN in media order, declared once with an alias. Each role line scopes one job in plain words ("only the brickwork and the window light") [proven], says what not to take and that it sets no camera position. A neutral plate's role line tells the model to drop the plate's flat light and light the shot from {the scene's source}; a plate whose angle the shot opens on gets that angle in words, and stays a reference. With a frame attached, name references by content. Tag only where the subject is visible [official guidance], or a stray tag pulls it into frame. If words and image disagree, the image wins.
Reference | Owns | Never |
Sheet and close-up | Identity, hair, wardrobe | Light, pose, layout |
Architecture, materials, light, exposure | Framing | |
Product photos | Shape, colour, label | Light |
Prop or state sheet | One prop's design; one visible state | Position; another state |
Staging map [best available, not yet validated] | Positions and facing, attached last | Being rendered in the shot |
Staging maps: see Still prompts, storyboard frames and diagrams → Staging maps, position images, scale sketches. People and products are lit by the location. Star, not chain: always attach the original anchors; a take's frame never becomes a reference (decided). The movement reference is a fallback only (see Planning, budgets and takes → Choosing each generation's method).
A labelled location diagram. Option: use it when the layout and distances must hold [single strong source]. Attach a labelled diagram of the place as an extra location anchor. Its role line says it gives the layout, positions and distances only, and that its lines, labels and colours are never rendered.
Two states of one prop. Option: use it when a prop changes inside a take [single strong source]. Attach two images of the one object, with a count lock ("exactly one candle, unlit and then burning").
Do it yourself
Attach the references in tag order (Higgsfield app)
The app's screens change often, so these steps say what to do, not where each control sits.
Add the images in the order your ACTIVE REFERENCES block declares them. The order the images sit among the job's attached references is the tag order: the first is @image1, the second @image2, and so on.
Add nothing the shot doesn't use: usually 2–6 images. The limits are 30 images (frames included), 10 videos and 10 audio clips, 50 in all.
Add an audio reference only with an image or a video beside it, and an end frame only with a start frame.
If you attach a start or end frame, the prompt names every reference by what it shows, with no @ numbers, so it doesn't matter which slot the frame takes.
Before you submit, read the attached references against the prompt: every tag points to an image that is there, and no image is left without a role line.