AI VIDEOS

Two clips, one room.

Generated shots arrive as strangers. Each one is a separate roll, and nothing in the model obliges the second clip to believe it is standing in the same room as the first. Continuity has to be written before it can be watched.

The hardest part of assembling generated video is not any single clip. It is the moment between two of them. Each generation is an independent roll, and nothing in the process obliges the second clip to believe it is standing in the same room as the first. The model has no memory of the set.

So continuity stops being something you repair in the edit and becomes something you write in the prompt. That is a small shift with large consequences, because it moves the work earlier, to the point where it still costs almost nothing. A mismatch caught in a sentence costs a sentence. The same mismatch caught in the timeline costs a regeneration and the take you had already fallen for.

Ask for the same character twice and you will often get two people who merely resemble each other. Ask for the same corridor twice and you will get two corridors that share a mood and disagree about everything else: the height of the ceiling, which side the windows sit on, whether it is morning. Nothing carries over unless you carry it.

This is not a defect worth complaining about. It is the shape of the tool. Once you accept that every generation starts from nothing, the job becomes obvious: decide in advance what has to persist, then say it every single time, in the same words.

The habit that changed my cuts most was reducing continuity to exactly three named constants, written once and pasted into every prompt in the sequence. Not ten. Three. A short list gets repeated faithfully, a long one gets paraphrased, and paraphrase is where continuity quietly dies.

Mine are usually the space, the light source, and one held detail. A narrow corridor with a green tiled floor. A single window camera left. A red coat. Those three phrases travel word for word between prompts, and everything else is allowed to vary, because everything else is what makes the shots feel like different angles rather than different films.

Describing light as a mood is the common error. Warm, moody, cinematic: those words produce a feeling, and a feeling can be produced two different ways in two clips. Describe the source instead. One window camera left. A practical lamp on the desk. Overhead fluorescents with no fill.

A named source is a physical fact the model can honour consistently, and it constrains shadow direction across the cut without your ever having to mention shadows. It also gives you something specific to check against afterwards, which a mood never does.

If a subject exits frame right, the next clip should open with them entering frame left, and you have to say so, because the model will not infer it. Screen direction is the oldest continuity tool there is, and it survives the jump to generated footage completely intact. Write the exit into one prompt and the entrance into the next.

The same goes for camera movement. A push that stops dead at the cut reads as an accident. A push that hands off to a slower push in the following shot reads as intent, and the two clips begin to feel like coverage of one moment rather than two separate attempts at it.

Sound is a generous liar. Music and room tone will paper over a continuity break so smoothly that you will not catch it until someone else does, usually in front of the client. So the last pass on any generated sequence is silent and full size, watching only for the things that changed when nothing should have. The silent pass finds in twenty seconds what three scored viewings will hide.

Two small habits sharpen that pass. Watch at full size rather than in a corner preview, because a break that is invisible at four hundred pixels is unmissable at full width. Then watch it a second time with your attention split differently: once following only the join between shots, once following only the held detail you named. Looking for one thing at a time finds more than looking carefully at everything.

The cut is where generated video either becomes a scene or stays a pile of clips, and nearly all of that is decided before a single frame renders. Name the space, name the light source, name one held detail, and carry the direction across the join. Then watch it silent and believe what you see.

THE NEXT WORKING NOTE

I am putting together the seam kit: the three-constant block I paste into every prompt in a sequence, a short list of light sources that hold across a cut, and the silent-pass checklist I run before anything reaches a client.

Want it when it ships? Reply with send me the seam kit and I will get it to you.

A QUESTION FOR YOU

Which continuity break keeps getting past you?

Reply with the specific one. The doorway that moves, the coat that changes shade, the light that swaps sides between two shots you had already approved.

If this was useful, forward it to someone assembling generated clips into an actual scene.