One generation with cuts, or three clips stitched

A three-shot sequence can now be produced two entirely different ways. You can write a shot list into one prompt and let a model that supports multiple prompts inside a single generation cut between the angles itself. Or you can generate three clips independently and assemble them on a timeline. Both end in a file of roughly the same length. They are not the same decision, and the thing that separates them is not quality — it's what happens when shot two is wrong.

What each path actually is

The single-pass route depends on the model exposing multi-prompt behaviour. In Versely's catalog that's a small, specific set: three endpoints carry a multi_prompt capability, all in the Kling V3 text-to-video line, including Kling Video V3 4K Text to Video. Fifteen video models carry a camera_control capability, which is a different thing — control over the move within one shot, not the ability to cut between shots. If you want the mechanics of prompting a shot list into one generation, single-pass multi-shot directing covers that ground and this post won't repeat it.

The stitched route is three ordinary generations plus an assembly step. In Versely that assembly is edit_video, an EDL render: one timeline that stitches, trims, speed-changes, captions and mixes into a single output. Clips have to be HTTPS URLs from your workspace or a prior generation. Merging videos is the task page for it.

Laid out on the axes that decide it:

	One generation with cuts	Three clips stitched
Continuity across the cut	Model-enforced. Wardrobe, set and light carry unless the prompt left them open	Nothing carries. You enforce it with references or you don't get it
Per-shot control	Low. One duration, one model, one resolution, one seed for all three	Total. Different model, length, ratio and reference set per shot
Cut timing	Whatever the model decides	Frame-accurate, and adjustable after the fact
Retake unit	The whole generation	One clip
Assembly cost	None	One editor export for the finished timeline
Fails as	A sequence that's 90% right and unusable	Three good shots that don't look related

Note the last row, because it's the honest summary. These two paths fail in opposite directions, and which failure you'd rather debug is most of the answer.

Continuity is the only thing single-pass genuinely buys

A model generating three shots in one pass holds the scene in one context. It knows the jacket, the room, the key light direction, and the actor's face are the same across the cut because it is producing all of them from one state. That is a real advantage and it is not reproducible by prompting three separate generations with the same paragraph — nothing carries between generations by default, and three independent runs will differ on every detail the prompt left unpinned.

The catch is that this advantage is strongest exactly where it's least often needed. Most three-shot sequences in short-form marketing are not continuous scenes; they're a hook, a demonstration and a product beat, frequently in different spaces. Nothing about them requires the jacket to match, because there is no jacket in two of the shots. Single-pass continuity is worth paying for when the three shots are the same moment from three angles. It's worth close to nothing when they're three different moments.

Where you do need continuity across separately generated clips, the fix is a reference anchor rather than a shared prompt — see reference stacking for what current endpoints accept, and consistency collapse for diagnosing which of the four kinds of drift you actually have.

Retake cost is where the decision is really made

Assume you accept a shot at some rate — call it p. You will not know your own p until you've run a few briefs, and it varies enormously by subject, but the arithmetic works with any value you plug in.

For three independent clips, each shot is its own trial. A rejected shot two costs one regeneration. Shots one and three are already banked and stay banked.

For one generation containing three shots, the unit of acceptance is the whole generation. All three have to land together, so the sequence succeeds at roughly p³. At p = 0.7 per shot, that's about 0.34 for the sequence — you're rerolling the majority of the time, and every reroll throws away two shots that were fine. Worse, the reroll re-rolls them: the shot you liked in attempt one is not preserved into attempt two, because it isn't a separable artifact.

This compounds with a second asymmetry. On the stitched path each shot can use a different model — a cheap fast endpoint for the establishing shot, a premium one for the hero beat. On the single-pass path every shot pays the hero rate, because there's one model and one resolution for the generation. Scene count economics has the per-scene math on multi-scene projects generally; the short version is that per-scene cost tracks the model and resolution you chose, not the number of scenes.

The threshold

Choose single-pass when all four of these are true:

1. Two cuts, not more. Three or more cuts in one generation compounds every continuity boundary, and the practical failure point tends to arrive well before a model's advertised ceiling.
2. The shots are one continuous space and time. Same room, same minute, same wardrobe. If a viewer would accept a hard scene change between two of your shots, you don't need the model to hold anything across it.
3. You're producing one variant, not a matrix. Single-pass output can't be recombined. If the brief is "three hooks and three end cards against the same middle shot," stitching gives you nine timelines from seven generations and single-pass gives you nine full generations.
4. The cut timing isn't load-bearing. If the cut has to land on a specific beat in the music or a specific word in the voiceover, you need a timeline, because the model chooses its own cut points.

Fail any one of those and stitch. In practice that means most ad work stitches and most narrative work — a conversation, a reveal, a reaction covered from two angles — is where single-pass earns its keep.

Assembling the stitched version without paying for guesses

The stitched path's assembly step is one edit_video call, and the workflow that keeps it cheap is to iterate at preview quality first. preview: true returns a free 480p render — subject to a short per-user cooldown, so it's a checking tool rather than something to hammer — and the final export at full resolution is charged once for the whole timeline regardless of how many clips are on it. That means cut timing, transition length and caption placement are all decisions you can make and remake before anything bills. Editor previews and final export has the cost scenario written out.

A workable order of operations:

1. Generate each shot independently, using whichever model suits that shot.
2. Reroll only the shots that failed, at the shot level.
3. Assemble the timeline in one edit_video call and preview at 480p.
4. Adjust cut points and transitions against the preview, not against a description.
5. Export once when the timing reads correctly.

FAQ

Does a multi-shot generation cost less than three separate generations?

It's one generation instead of three, but it's billed on the same basis as any other clip from that model — length and resolution — so a single generation long enough to hold three shots is not automatically cheaper than three short ones. Ask the agent for an estimate_cost on both shapes before assuming either way.

Can I take a good shot out of a multi-shot generation and reuse it?

You can trim it out with cut_video and drop it on a timeline, so yes, mechanically. What you can't do is regenerate the other two shots around it — the next attempt is a fresh generation with its own version of the shot you kept. Treat multi-shot output as one artifact for planning purposes even though you can slice it afterwards.

If I stitch, how do I stop the three clips looking unrelated?

Anchor them to the same pixels rather than the same paragraph. A shared reference image set, or a named reusable asset carried across scenes, does what a repeated description doesn't. Named assets covers registering a character or product once and referencing it by key in every scene.

Is the cut quality worse when I stitch?

The cut itself is cleaner when you stitch — you choose the exact frame, and you can adjust it after seeing it. What's worse is everything either side of the cut agreeing with itself, which is the continuity problem above. Sharp cut, mismatched contents is the stitched failure; smooth transition, unusable third shot is the single-pass one.