On AI video continuity
You wrote a shot list. You got one long drifting take. This is what is actually going wrong — the cut, the character who changes across it, and the two people who never quite look at each other — and how far better wording gets you before you need something other than words.
You wrote it clearly. "Wide shot of the cafe. Cut to a close-up of her face. Cut to his table. Cut back to her as she smiles." Four shots, no ambiguity. What came back was a single unbroken push-in that started wide and ended somewhere near her shoulder, drifting the whole way, and by the time it got there she had a different nose.
Everyone who has tried to make a scene rather than a clip has hit this. It reads as the model ignoring you, and most advice treats it that way — write the prompt harder, add "hard cut", put the timings in capitals. That advice caps out quickly, because the problem is not emphasis. Three separate things are going wrong, and only one of them is about wording at all.
A video model produces a single run of frames in one pass, and every frame is conditioned on being plausibly adjacent to the last. Smoothness is not a stylistic preference it happens to have — it is close to what the thing is built to do. A cut is the one event in film grammar that is deliberately not plausible frame-to-frame: two adjacent frames that share almost no pixels and yet belong to the same scene.
So when you write "cut to a close-up", the model has two readings available. One is a discontinuity. The other is a description of a camera that ends up close on her face. The second reading is enormously more probable under everything it learned, so it dollies in. It is not ignoring the word cut. It is resolving it the way that costs it least.
A cut is the one event in film grammar that is deliberately not plausible frame-to-frame. Smoothness is what the model is built for.
Suppose you do get four distinct shots — by generating them separately, which is what most people fall back to. Now the second failure arrives. Her coat changes colour. The cafe has a different window. He was on her left and is now on her right.
Each generation resamples everything you did not pin down, and prose pins down far less than it feels like it does. Reference photographs help, but they are specific about what they carry: a reference image fixes what a face looks like. It says nothing about where that person is standing, which way they are turned, or where the camera was when you saw them. Identity and geometry are different problems, and stills only solve the first.
This is the failure that makes an otherwise good result feel amateur without anyone being able to say why. Two people are having a conversation. You cut between them. Both of them are looking slightly frame-left. An audience reads that instantly and wrongly: not as two people facing each other, but as two people both watching a third person off-screen.
The rule underneath it is old and mechanical. When you cut between two people in a conversation they must look in opposing screen directions — she is frame-left looking right, he is frame-right looking left. Generate each shot in isolation and there is nothing forcing that opposition, so roughly half the time you get a matched pair and half the time you do not.
Its sibling is the character who stares down the lens. That one usually is not a model failure at all — it is a framing mistake. If you put the camera on the eyeline, directly behind the person she is supposed to be looking at, then looking at him is looking into the lens, and the model is being obedient. Offsetting the camera forty to sixty degrees off that axis sends the look across the frame, which is what you actually wanted and what the shot was missing.
Within the limits above, some shot lists work far better than others. Three things carry weight.
Name spans, not transitions. "Then cut to" is a description of an edit. "4.0–7.5s" is a constraint you can check. Give every span a shot size and hold it.
Say which way people face, in screen terms. Frame-left and frame-right are unambiguous in a way that "facing each other" is not, because the second one is only true relative to a camera position you have not specified.
Describe faces physically, never by emotion. "Devastated" has no pixels in it. "The eyes drop to the table a beat too long before she answers"does. Emotion labels also push generators towards the extremes of the label, which is how you end up with someone weeping when you asked for restraint.
0.0–4.0s WIDE. The room, both of them in it. Camera static.
She is frame-left, he is frame-right, a table between them.
4.0–7.5s CU on her, camera 50° off the eyeline, slightly low.
She is frame-left looking RIGHT, across the frame, at him.
The smile arrives late and does not reach the eyes.
7.5–11.0s CU on him, reverse. Frame-right looking LEFT.
He holds the look without blinking. One swallow, no words.
11.0–15.0s WIDE again, same framing as the opening.
Neither of them has moved. That is the point.That will get you a materially better result than a paragraph of prose. It will still, fairly often, hand you back one continuous move — because every line of it is a description of a cut, and the model has to choose to believe you.
The step-change is not a better sentence. It is giving the model a reference clip that genuinely cuts. Then the discontinuity is in the pixels rather than in your prose, the timings are visible rather than asserted, and the framing of each shot is something the model can see instead of infer. A motion reference and a still reference carry different things, and this is the split worth remembering: stills carry identity, a clip carries geometry and timing. Most scenes need both.
Doing that by hand means owning a camera and the scene you want to shoot, which defeats the point. The workable version is to block the scene in a rough 3D view — grey boxes for the room, placeholder figures for the people, a real camera you position per shot — then record that playback and hand it over. Nothing in it needs to look good. It only needs to be geometrically true: where the bodies are, which way they face, where the camera sits, and when it changes.
Stills carry identity. A clip carries geometry and timing. The greybox does not need to look like anything — it only needs to be true.
The whole loop in 65 seconds: a cafe blocked in grey, four camera setups, the cuts surviving into the generated video.
Blocking it once also disposes of the eyeline problem structurally rather than by wording. If both characters are objects in one scene, the direction each of them faces is a single fact that every shot inherits. Put the cameras on opposite sides of the line and the screen directions come out opposed automatically, because the geometry says so. You stop writing "she looks at him" and hoping, and start pointing a camera at a person who is already looking the right way.
Honest limits. Clips run to about fifteen seconds, which is the ceiling on the current video models, not a product decision — enough for four or five shots and one complete beat, not a scene of dialogue. Faces still drift under hard motion. A reference clip constrains framing and timing well and lighting barely at all. And none of this makes a badly chosen shot a good one: blocking gives you control over the cut, which is a different thing from knowing where to put it.
But the specific failure at the top of this page — the shot list that came back as one drifting push-in — is not a limitation of the models. It is a mismatch between what a cut is and what a sentence can be. Stop describing the cut. Show it one.
Because a generation is one continuous sample. The model reads "then cut to a close-up" as a description of camera behaviour rather than an instruction to break continuity, and the smoothest way to satisfy a description of two framings is to move from one to the other. So it dollies in instead of cutting.
Because nothing in the prompt asserts that the two shots are the same room and the same person. Each shot is resampled, and anything you did not pin down is re-guessed. Reference photographs pin the face; they do not pin where the person is standing or which way they are facing.
Usually because the camera was put on the eyeline. If the person they are supposed to be looking at sits directly behind the lens, then a correct look is a look into the lens. Moving the camera 40 to 60 degrees off that axis sends the look across the frame instead.
Screen direction. When you cut between two people they have to look in opposing directions — one frame-left looking right, the other frame-right looking left. Generate each shot without a shared geometry and both tend to come back looking the same way, which reads as two people addressing a third person off-screen.
It helps more than prose does. Naming spans — 0 to 4 seconds, 4 to 7.5 seconds — and giving each one a shot size and a screen direction gives the model something checkable. It does not solve the problem, because timings are still a description of a cut rather than a cut.
Showing it a cut. A reference clip that is genuinely discontinuous carries the cut in its pixels instead of in your prose, and the model follows the framing and the timing it can see. That is what a blocking tool is for: compose the shots, record the result, and hand that over as the motion reference.