How a shot is made

A start frame is the exception, never the method.

Seedance 2.5 is the one video model in this stack. These pages take a shot from approved anchors to a keeper, the take you choose for the film. Camera, light, performance and speaking shots are taught in their own chapters: Cinematography, Performance and Voice and lip sync.

The default [proven]: Seedance 2.5 in omni-reference mode (the mode that takes your approved anchors as references beside the prompt), the approved anchors attached as image references (character sheets, identity close-ups, location plates, product photos, prop sheets), and a prompt that says exactly what happens, for one shot or several. A start or end frame is added only where the plan names a start-frame case (see Video with Seedance → The six start-frame cases).

No production video before the client approves the anchors. Until then, the only video you make is the face motion test of each candidate face and the place motion check of each candidate place (see Writing the treatment → Face motion tests and Photoreal anchors (Nano Banana Pro) → The place motion check). The two are the same kind of step: an internal check of one candidate asset, a face or a place, not of a method, made before the client sees it.

Every generation

1. Read the scene, its row in your shot plan (the plan you made before generating; see Planning, budgets and takes) and your take log. The storyboard shows intent only; it is never attached or matched.

2. Choose the shape (below) and assemble the references (see Video with Seedance → Attaching references).

3. Write the prompt from a template: A, B or D for shots where nobody speaks a principal line (see Video with Seedance → Templates A, B and D); C, E or F for speaking and singing shots (see Speaking shots and lip sync → Templates C, E and F).

4. Validate and pre-flight the prompt (see Video with Seedance → Validate, then pre-flight).

5. Get the estimate approved before any paid job (see Planning, budgets and takes → Costs and the estimate).

6. Submit drafts in the generation rhythm. First, make one sample shot per scene, two 480p drafts each, and review all the samples together. Then make the remaining shots: each gets one new take in each batch, and no shot gets another take while its problem is unresolved (see Planning, budgets and takes → The generation rhythm).

7. Finish the keeper at 1080p, review the 1080p finish at full size and log it.

The shape

Choose one of three shapes for each generation [proven for the first two; separate cases from mixed sources]:

Make separate generations for:

• a packshot;

• a change of world that isn't designed as a transition;

• a speaking shot whose recorded line must stay exact (one speaker per generation);

• more than three acting characters without a staging map or previz (a rough 3D move made first to fix the path);

• contact that must read;

• a dialogue exchange longer than one generation holds;

• heavy physics beside close micro-performance;

• one moment from two vantages.

A shot that keeps failing becomes a separate generation too. Retake it as a fragment 2–3 shots long, with the same header and geometry, opening mid-situation.

Why: the placing films at Higgsfield's 2026 festival worked this way. The Field Agent built a 30-second, eight-shot card from eight reference images and no start frame, each reference limited to one job; Listen to Meh made its 30-second one-take from references alone and kept a start frame for one close-up.

Sources

• Higgsfield Global Film Festival placers: The Field Agent by Zack Murray (project); Listen to Meh by @nue_nue (project); PJ Accetturo (newsletter).