Continuity between shots — continuous shots, cast, look and seed in Cinema

The hardest thing about a generated film is not any one shot. It is the join between two shots: the coat that changes color, the room that gets a second window, the face that drifts. Every engine renders a clip in isolation, so continuity has to be built around the engine. Cinema stacks four layers to do it, and this guide is about what each one holds and where it stops.

Layer one: continuous shots

The strongest join is the one the engine cannot get wrong: start the next scene on the last frame of the previous one.

With continuous shots on — it is on by default, in the Look panel — Cinema renders scene N+1 from the last frame of scene N's picked take. Concretely:

1. When scene N has a take, the studio grabs its final frame right in the browser.
2. The frame is uploaded to your storage, in a Film_frames folder named after the film.
3. It is sent as the start frame of scene N+1 — the first reference the engine sees.

A scene whose predecessor is still rendering shows waiting and starts as soon as the frame exists. If the previous take's link has expired, the studio re-signs it and tries again; if the frame genuinely cannot be grabbed, the scene renders fresh and says so in a note, rather than failing.

What this buys you: the light, the set and the framing at the cut are identical, because they are the same pixels. What it does not buy you: a cut. Continuous shots make a long take, not a montage. When you want a hard cut — new location, new angle — turn the toggle off for that film, or write the new scene so the first second is a move away from the frame (a whip pan, a push through a door).

Two rules keep the chain honest:

• Re-render scene N, or pick another of its takes, and scene N+1 is marked stale. Its take opened on a frame that no longer leads the cut. Re-render it or accept the join.
• The chain is inert where it cannot work: engines with no start-frame control, and the studio's private mode. The Look panel says which.

Price: a start frame bills as one frame on engines that price frames, and zero references — the same arithmetic the server uses.

Layer two: the cast

Continuous shots hold the set. They do not hold a character who walks into frame in scene three. For that, the scene's cast rides as references.

Add characters to a scene's cast (the Cast picker in the Storyboard's scene panel, or @mention them in the action). Each cast member's sheet — the reference picture from their character bible — is sent with the scene, and the engine prompt binds the name to the picture: "Mina (reference image 1)". Speakers of a scene's lines join its cast automatically; on engines that accept an audio reference (Seedance 2.x, Hailuo 3), the first voiced speaker's sample goes too, so the line comes out in their voice rather than one the engine invents. One catch: a continuous shot opens on a frame and sends no reference pictures or voice sample — its cast carries over from the frame it opens on, and its voices ride as descriptors.

What a sheet holds: face, hair, build, the outfit on the sheet. What it does not: a change of outfit mid-film (make a second sheet), and — this is the important caveat — engines that read no references. Sora 2 Pro, Runway Gen-4.5, Grok Imagine, Veo 3.1 Lite and Hailuo 3 Max render from the prompt alone; on those the cast is redrawn every scene, and the storyboard warns you. If your film is about a face, pick an engine that reads references: the Seedance family, Kling 3.0, Hailuo 3, Veo 3.1, Wan 3.0. The character consistency guide compares them.

A character's persona (their personality, their speech) never rides a video prompt — only their age, gender and look. The engine renders a picture; it does not need to know how she argues.

Layer three: the look

Continuous shots and cast hold content. The look holds style: one engine, one frame format, one resolution, and the director's rack — film type, era, tempo, camera gear, lens, aperture, palette, lighting, plus a negative prompt. Set once, folded into every scene server-side as labelled sentences the engine reads the same way each time.

The look is why scene one and scene twelve cut together even when nothing else connects them. A film noir with an anamorphic lens at f/1.4 under practicals looks like that on every card; nobody has to remember to write "neon noir" in scene nine. The director's rack guide explains the knobs.

Do not change the look mid-film unless you mean it. Every scene rendered before the change now differs from every scene after, and Cinema will not re-render them for you — it marks nothing stale, because a look change is a decision, not an accident.

Layer four: the seed

A seed pins the engine's randomness. In the Look panel, one seed applies to every scene of the film on engines that honor it. Same seed, same prompt, same engine → the same clip, which is what you want when you fix a typo in scene four and re-render only that.

It is also the weakest layer. A seed does not make two different prompts render alike; it makes the same prompt render alike. Use it for reproducibility, not for continuity. The seeds and remix guide covers the arithmetic.

Putting the four together

A practical order, for a three-scene film:

1. Write the three scenes in the Script, cast in place, dialogue inside the scene it belongs to.
2. Set the look. Pick an engine that reads references if the cast matters. Leave continuous shots on.
3. Render scene one alone — one to four takes. Pick the best. This take's last frame is now the film's spine.
4. Render scenes two and three. They chain. If two is right and three is not, re-render three; two does not move and you do not pay for it again.
5. Cut, add sound and titles, export.

Where continuity still breaks — a hand that changes, a prop that vanishes — the fix is the same as in any edit: cut around it. Trim the take, crossfade, or pick another take. The engine will not be perfect; the joins can be.

FAQ

Does continuous shots work with every engine?

It needs an engine with a start-frame control. Where the engine has none, the toggle is inert and the Look panel says so. Motion One and the Seedance, Kling, Veo 3.1 and Hailuo 3 engines support it.

Why is scene two "stale"?

Because scene one changed after scene two was rendered — a re-render or a different picked take. Scene two's take opened on a frame that is no longer the end of scene one. Re-render it, or keep it if the join still looks right.

Can a character keep the same voice across scenes?

On engines that take an audio reference (Seedance 2.x, Hailuo 3), yes — the character's voice sample goes with every scene where they speak. Elsewhere, the prompt carries a one-line voice descriptor and the engine does its best.

Is the start frame charged as a reference?

No. It is one frame, on engines that price frames, and it replaces references rather than adding to them.

Open a film in Cinema and render scene one — the chain starts from there.