Choosing each generation's method

References only is the default. Every other method needs a named case and a one-line reason.

The method column of your shot plan names how each generation is made. A face alone is never a reason for a frame.

The cases themselves are in Video with Seedance → The six start-frame cases. Other modes, where they apply: a previsualisation (previz) video reference, extension, video-to-video, and a code render for motion graphics. Previz suits single-subject shots that need an exact camera speed, and is weak on several subjects crossing in shared space [best available, not yet validated]. Rely on it in a plan only once the method has worked for you; until then, write the camera move in full with a staging map (see Camera and lens → Exact paths).

Frames, when a method names one:

• Make them only after the client approves the anchors, fresh from the approved sheets (see Photoreal anchors (Nano Banana Pro) → Internal start and end frames).

• Read a storyboard frame, the client's or your own, and write its framing, pose and prop handling into the prompt in words; it is never attached (decided). A take screenshot may join an image generation only as a composition and continuity guide; its role line takes nothing else from it.

• A start frame pins only the opening of a multi-shot clip.

• No end frame without a start frame: the platform takes an end frame only beside a start frame. Plan a shot that needs only an exact ending as references only, with the landing in words and a 1–2 s tail handle. Where the ending must be one exact client photograph, it is the exact-landing case with both frames, the start made fresh from the anchors.

The movement reference is a fallback, never planned per shot (decided). It is a self-filmed phone reference: timing, body weight and expression only. The look comes from the anchors, the filmed person's look is excluded in writing, and consent is signed. It may join a silent hero acting beat only after that shot has failed after several drafts, and only when you choose it in that batch review. No plan row names it in advance.

The 360° room turn (decided) is an optional extra place draft, priced like one: from one fixed spot in the place, the camera turns slowly through a full circle, made from the approved master plate. Read its frames every 0.5 s to learn what the room holds behind the master plate's camera, and write that into the prompts of the reverse and coverage plates. What it shows only guides the written prompt: no frame from it is ever attached to any generation, and it never becomes an anchor.

Visual effects (VFX): generated or composited [best available, not yet validated]. Decide it per shot, with who composites it and when it enters:

• Generated in the shot: an effect with a physical cause one world can hold (rain, grinder sparks, steam, a crack from pressure, a creature's fixed rule), written as cause and effect (see Video with Seedance → Physics).

• Composited: anything that must be exact, readable or identical across shots (screen UI, a visualised signal, logos, text), a transformation that must land on an exact frame, and any element over a live plate. Flat placement on a held frame is a graphics job (After Effects, or Fusion in DaVinci Resolve); tracking, keys and roto are finishing.

• Element plates: an element generated alone, on plain black for light that adds (sparks, fire, light streaks) or on flat mid-grey for an element to be keyed. The camera matches the host shot's move or is locked. Seedance gives no alpha, so the plate is screened or keyed in finishing. Each is its own low-risk generation.

• Creatures and transformations: the creature's sheet variants. A transformation gets start, during and end state sheets, and is made inside one generation only as the change-of-view case, with a fallback cut.

Sources

• BALLAST by @studio141 (project) and Lullaby by Keigo Matsumaru (project) (previz scope); PJ Accetturo (newsletter) (frames by exception).