Six Shots, One Prompt

This is the archive post in the Model Notes series, and it is deliberately out of sequence. Everything else in this run shipped in the last eight weeks. Kling 3.0 shipped in February, which in this industry is a geological era ago.

We are writing it anyway, and putting it at the foot of the series, because if you read the seven posts above this one you will notice something. Every major release of this quarter is arguing with a decision Kling made in February. Native multi-shot. Audio generated with the picture rather than after it. A storyboard as a first-class input. Those ideas were not obvious a year ago. They are now the assumed baseline, and it is worth remembering who assumed them first.

Then we are going to argue with it ourselves, because the release also contains a claim we think is wrong, and it is in the headline.

What actually shipped

Kuaishou launched the Kling 3.0 family on 5 February 2026, comprising four models rather than one.

• Video 3.0
• Video 3.0 Omni
• Image 3.0
• Image 3.0 Omni

Video. Up to 15 seconds duration. Kuaishou’s own release emphasises photorealistic output with lifelike characters in expressive, dynamic performances. Coverage at launch also described native 4K output at high frame rates and a unified multi-modal architecture processing everything through a single pipeline rather than chaining separate tools. Those specifics came from secondary reporting rather than the Kuaishou release, and we report them as reported.

Image. 2K and 4K ultra-high-definition output.

Multi-shot storyboarding, the headline feature. Video 3.0 Omni introduced a storyboard input in which you specify, per shot: duration, shot size, perspective, narrative content and camera movement, with spatial continuity maintained automatically between them.

Omni Native Audio. Generated across English, Chinese, Japanese, Korean and Spanish, with multiple accents and dialects, supporting multi-character dialogue scenes and precise control over delivery and speaking order.

Motion and camera control. Intelligent camera-angle adjustment covering classic shot-reverse-shot dialogue through to cross-cutting dialogue and voice-over.

Availability at launch. Exclusive early access for Ultra subscribers, with public availability following.

Scale, as stated by Kuaishou. More than 60 million creators, more than 600 million videos generated, and more than 30,000 enterprise clients.

Why it still matters seven months later

It moved the storyboard inside the model. Before February, multi-shot meant generating shots separately and hoping they would cut together. Kling made the sequence the unit of generation. Read back through this series and watch the idea propagate: Gemini Omni’s first-and-last-frame control in August, FLUX 3’s keyframe-to-video in July, MiniMax’s native multi-shot modelling. Everyone arrived at the same conclusion, which is that prose is a terrible way to specify a shot and a diagram is a good one. Kling got there first.

It named shot size and perspective as parameters. This is the part we find quietly moving. Somewhere in the Kling 3.0 spec there is a field called shot size, with values that are the same values a first AD has written on a call sheet for a century. Wide. Medium. Close. Not “cinematic.” Not “epic.” The actual vocabulary of the craft, promoted from adjective to parameter. That is a design team that talked to filmmakers.

It treated shot-reverse-shot as a thing worth building for. Shot-reverse-shot is the most common construction in narrative cinema and the single hardest thing for a generative model to keep coherent, because it requires the model to understand that there is a line and that both shots are on the same side of it. A release that named it explicitly was a release taking dialogue scenes seriously at a moment when the rest of the field was still making beautiful, silent, single-shot moodscapes.

Audio in five languages with dialect control, in February. Native dialogue audio became a standard expectation over the following two quarters. It was not standard when Kling shipped it.

The argument we want to have

Kuaishou’s own release headline says Kling 3.0 ushers in “an era where everyone can be a director.”

We want to disagree with that sentence carefully, because we agree with the half of it that matters and we think the other half does real damage.

The half that is right, and it is the bigger half. The barrier to filmmaking has never been talent. It has been access: to equipment, to a crew, to the room where the budget gets approved, to a film school that costs more than a family earns in three years. There is no shortage of people who can see a film in their head. There has only ever been a shortage of people permitted to make one. Any tool that widens that door is doing something genuinely important, and we will defend that against anyone who finds it distasteful. The gatekeeping was never protecting quality. It was protecting gatekeepers.

The half that is wrong. “Everyone can be a director” collapses two different things: being able to generate and knowing what to generate. The first is now essentially solved. The second is the entire job.

A director is not someone who can produce images. A director is someone who knows, when the six shots come back, which one is lying. Who can say the performance is good and the shot is wrong, and here is why. Who understands that the reverse does not cut because the light moved, and that the light moving is a bigger problem than the resolution. Who knows what the audience should be feeling at second 28 and has built every decision backwards from that.

None of that arrived in February. None of it is in any release in this series. It is learned the way it has always been learned: by watching enormous quantities of work, by making bad things and understanding why they are bad, by developing an eye slowly and then trusting it.

So everyone can now make a film. That is true, and it is wonderful, and it is the most democratising thing to happen to this craft in a hundred years. Everyone can be a director is a different claim, and the gap between the two is not a gap in tooling. It is a gap in looking.

The grammar test

For multi-shot models we run the line test, and it remains the hardest test in this series.

Build a six-shot storyboard of a two-person dialogue scene.

1. Wide, both characters, establishing the geography.
2. Medium, Character A.
3. Medium, Character B, reverse.
4. Close, Character A.
5. Close, Character B, reverse.
6. Wide, return to the establishing geography.

Generate it as one storyboard. Then check three things.

1. Is the line respected? Does A consistently look frame-right and B frame-left across every reverse? A single crossing breaks the scene and no grade fixes it.
2. Does the key light stay on the same side of the room? Not the same side of the face, but the same side of the room. In a correctly lit reverse, the key comes from the same physical source, which means it lands on opposite sides of the two faces. Most models light both faces identically and it looks subtly, unplaceably wrong.
3. Does shot 6 match shot 1? The return to the establishing wide is the continuity audit. If the room has changed, the sequence was never coherent. You just could not see it until you came back.

Kling 3.0 was the first model we tested that passed test one with any reliability, which in February was remarkable. Test two remains inconsistent across the field, including here. Test three is the honest measure of whether a multi-shot system is modelling a space or generating six plausible pictures, and it is where every model in this series still has work to do.

If you run only one test from this entire series, run this one. It is the closest thing we have to a real examination of whether a model understands cinema or merely resembles it.

Where it breaks

• Fifteen seconds was mid-field in February and is short now.
• Light across reverses still does not hold reliably. Specify the source position explicitly in every shot of a storyboard.
• Generation time for a full multi-shot pass is slow enough that iteration hurts. Block elsewhere, finish here.
• Six shots is a scene fragment rather than a scene. Useful, not sufficient.
• The 4K and architecture specifics come from secondary reporting rather than the Kuaishou release. We have flagged them as reported and would treat them as such.
• Prompt budget. Working practice on Kling 3.0 puts a real ceiling on prompt length, so plan on structured compression rather than prose. Six shots of specification inside a tight character budget is its own discipline.

Inside Pindow

Canvas and the 3D Stage, and Kling 3.0 is the reason the 3D Stage exists as its own step in our pipeline rather than as a checkbox.

The moment a model accepts a six-shot storyboard with shot size, perspective and camera movement as structured fields, the bottleneck stops being the model and becomes the storyboard. Where does the camera actually go? Which side of the line is it on? Where is the key? Those are spatial questions, and answering them in prose is how you get a scene that crosses the line without anyone noticing until the edit.

So block it in 3D, export the geometry as structured shot specifications, and feed the model a storyboard that is spatially true rather than verbally described. The 3D Stage is not a rendering tool. It is a decision-recording tool, and it exists so that the line, the eyeline and the key light are facts in the project rather than hopes in a sentence.

The prompt budget is the other half of our job here. Six shots of specification against a tight character ceiling is exactly the problem the Prompt Engine was built for: structured compression that preserves the load-bearing craft language of shot size, lens, light source and direction, and movement, while discarding the decorative adjectives that consume budget and add nothing. Most prompts are sixty percent atmosphere words doing no work.

Working notes

Write the storyboard before you write the prompt. Six shots, with shot size and camera position named for each. If you cannot fill that grid, you do not have a scene yet.

Name the light source once, per scene, and then repeat it in every shot. Not “the same lighting” but the actual source, temperature and position. The model has no memory of your intent between shots.

Draw the line. Literally. Somewhere in your notes there should be a mark showing which side of the axis the camera lives on. Then check every reverse against it.

Always return to the establishing wide. Even if you cut it later. It is the cheapest continuity audit available.

Budget your characters like film stock. Craft nouns first. Atmosphere adjectives last, and only if there is room.

Seven months on, the field has caught up to February on almost every specification and on none of the ideas. Longer clips, higher resolutions, more references, faster generation. All of it real, all of it welcome, and none of it addressing the question the line test asks.

The tools change. The grammar doesn’t.