Single-Pass Multi-Shot: Directing Cuts Inside One Generation

For as long as text-to-video has existed, a generation has meant a shot: one camera position, one continuous take, cut together with other generations afterward if you wanted a sequence. That's starting to change at the model level. A handful of frontier video models now take a prompt describing more than one shot — a wide, then a cut to a close-up, then a cut to a reverse angle — and return a single generation that contains all of them, transitions included. The prompt stopped being a scene description and started being something closer to a shot list.

What's actually shipping

Black Forest Labs is explicit about this in its own release notes: FLUX 3 Video can "create multiple scenes and camera angles within a single video, while keeping the sequence coherent." The same release describes a second, related capability — starting from an image, specifying an end frame, or setting multiple keyframes in a clip, with the model connecting them in sequence rather than requiring a separate generation per segment.

Kuaishou's own announcement for Kling 3.0 describes a more structured version of the same idea: up to six distinct camera cuts defined within a single prompt, with the model handling shot composition, transitions, and subject continuity across all six automatically, rather than each cut being planned by the person prompting it. That's a materially different framing from FLUX 3 Video's "keep the sequence coherent" — Kling's version is closer to a storyboard the model executes than a scene the model interprets.

Both are describing the same underlying shift: the atomic unit of a generation used to be a shot, and for a growing set of models it's now a sequence of shots.

What's live in Versely's catalog today

Versely's own listings for these model families are worth being precise about, because "same version number" doesn't automatically mean "identical feature set" once you're comparing a lab's flagship announcement to what's actually deployed and billable.

Flux 3 Text to Video and Flux 3 Image to Video are the same FLUX.3 model family Black Forest Labs describes above — the multi-scene, multi-angle capability isn't a separate toggle in BFL's own material, it's a property of how the base model follows a prompt. That means it's worth testing directly with a multi-shot prompt structure rather than assuming a special parameter exists for it; there isn't one to flip, because BFL doesn't describe it as one.

Kling Video V3 Pro Text to Video and Kling Video V3 Standard Text to Video carry a multi_prompt capability in Versely's own catalog data — the closest confirmed match to the storyboard behavior Kuaishou describes. What Versely's listing doesn't independently confirm is the specific six-cut ceiling from the press release; treat that number as what Kling's own announcement states, not as a guarantee reproduced on every endpoint. Kling 3 Turbo, the faster sibling line, does not carry the same multi-prompt tag in Versely's catalog — the turbo tier appears to trade that capability for speed, so it's not the one to reach for if directing multiple cuts is the point.

Whichever endpoint you use, the full current video roster is on /models, and /best/best-text-to-video-model ranks the text-to-video field if multi-shot isn't the only thing you're optimizing for.

Why the third cut is usually where it fails

Here's the practical pattern worth planning around: a two-cut generation — a wide followed by one clean cut to a close-up — tends to hold together. Ask for a third distinct cut in the same generation and coherence gets noticeably harder to trust, even on models that explicitly advertise a higher ceiling than three.

The mechanism is the same one that makes temporal consistency hard within a single continuous shot, just operating across a cut boundary instead of across frames. A model that hasn't been explicitly told what a background prop, a character's exact wardrobe detail, or a light source looks like re-decides those unpinned details every time it has room to — normally that shows up as flicker within a shot. Across a cut, the same gap shows up as full-blown inconsistency: the jacket is a different shade, the room has a different window, the actor's hair is parted the other way. Two cuts give the model one boundary to hold consistent. Three cuts give it two boundaries, and the second one is compounding on whatever drift the first one already introduced — which is exactly the failure curve you'd expect from a model re-deciding anything the prompt left open.

The practical mitigation isn't "wait for a better model" — it's treating a stated shot ceiling as a maximum to test cautiously toward, not a target to prompt for by default. Two clean, well-specified cuts you can trust beat four ambitious ones you have to regenerate.

A prompting structure that actually directs

The instinct to write one flowing paragraph describing a whole sequence undersells what these models can take. A structure closer to an actual shot list gives the model explicit boundaries to hold onto:

SHOT 1 (wide): [subject], [setting], [camera position]. [Lock any detail that must carry into shot 2 — wardrobe, prop, lighting direction.]
SHOT 2 (cut to close-up): Same [subject], same [wardrobe/prop from shot 1]. [New framing/angle.]

Repeating the locked details — the exact wardrobe, the exact prop, the light direction — at the start of each subsequent shot rather than assuming the model remembers them from shot 1 is the single highest-leverage habit here. It's redundant on the page and it's what keeps the second cut from re-deciding what the first one already established.

Walkthrough: a two-cut generation in Versely

1. Write shot 1 as a complete, standalone description — subject, setting, camera position, and every detail (wardrobe, key prop, light source) that has to survive into the next cut.
2. Write shot 2 starting with an explicit restatement of those locked details, then the new framing or angle for the cut.
3. Generate on Kling Video V3 Pro Text to Video if the multi-prompt structure above is the point, or Flux 3 Text to Video if you want to test the same technique on BFL's model.
4. Review the cut boundary specifically, not just the overall clip — pause on the frame just before and just after the transition and check the locked details against each other.
5. If shot 2 drifted, the fix is almost always to over-specify the locked details further, not to rewrite the whole prompt. Add a phrase like "unchanged from the previous shot" directly after the detail that drifted.
6. Before adding a third cut, confirm the two-cut version is solid. If it isn't, a third cut inherits and compounds whatever gap the first two didn't close.

If the goal is closer to a full scene sequence than a two- or three-cut clip, compare the multi-shot approach against building the same sequence as separate scene-by-scene generations stitched together afterward — for anything past three or four distinct cuts, that older approach is still the more reliable one, even as the single-pass ceiling keeps climbing on the model side.

FAQ

Is a single-pass multi-shot generation cheaper than generating each cut separately? It's one generation instead of two or three, which is fewer separate renders to manage, but it's still billed on the same per-generation basis as any other clip from that model — check the model's own listing on /models before assuming a multi-shot prompt is automatically the more economical path for a given sequence.

Can I mix this with reference images to lock a character across the cuts? On models that accept both a text prompt and reference images in the same call, yes — and it's a stronger anchor than repeating a wardrobe description in text alone, for the same reason a reference photo beats a text description anywhere else in these catalogs.

Should I always ask for the model's maximum stated shot count? No. Treat the advertised ceiling as evidence of what the model is built to attempt, not a target for a real brief. Two shots you trust outperform five you have to regenerate, and the practical failure point tends to show up well before most models' stated maximum.