All essays

    A good shot still has to make the cut

    Watch two generated shots back to back and the problems become harder to ignore. A look at continuity, rhythm and knowing when to stop making more takes.

    Ricky walking across the savanna in a five-second AI motion study
    From our studio development work, shown here to illustrate the process.

    What happens after the cut?

    It's easy to fall for a generated shot when we watch it on its own. Then we put it next to another one. The character seems to be looking at a different door. The light has switched sides. A hand begins a movement it already finished before the cut. Each image may be convincing, but something between them has broken. We only find that out when we start editing.

    A shot needs a reason to be there. It might tell us something new, pull us towards another character, keep us waiting or let us breathe for a moment. Writing that reason down can help before requesting another version. If two shots do the same thing, one may be unnecessary. Losing a favourite image hurts, especially after spending time on it. Sometimes that's exactly what the scene needs.

    The coat matches. Does the character?

    Faces, clothes and objects are often the first things we check. Next comes the space: which way someone is moving, where the doorway is, who is standing opposite whom. But we also need to follow what's happening to the character. What does she know by this point? What is she afraid of? Does this reaction belong to the moment before it? The coat can match perfectly while the person wearing it gets lost between shots.

    Runway describes using image references to carry characters, objects and environments into new images. Its documentation also notes that adding variables opens up more room for variation. A reference gives us a useful starting point. We still need to compare each result with the person and scene we're trying to create, rather than assume the tool has kept everything consistent.Runway: creating with image references

    Choose the take that works in the edit

    Give the takes clear names and view them at the same size. Watch the whole movement, then try each one between its neighbouring shots. The less spectacular take may get us into the next action more cleanly. Leave a short note about why it was chosen. Otherwise, the next round of revisions can easily remove the detail that made it work.

    That's the kind of work our AI Video Machine is built around: a shot, several takes, notes for a revision and a check before export. The chosen version stays with the reasons behind the choice. The example on this site shows that structure; it doesn't document a client production or present measured savings or performance results.

    What exactly needs fixing?

    'Make the shot better' leaves a lot to guess. 'Keep the character, framing and light, and let the hand finish moving before the cut' gives us something to check. Saying what must stay matters as much as naming the problem. It protects the parts that already work, and gives the next review a clear question: did this version fix what we asked it to fix?

    Sound deserves its own pass. Can we understand the dialogue? Does it arrive at the right moment? Does it connect with the shots around it? Google DeepMind identifies natural, consistent speech and audio synchronization as areas still being refined in Veo. The tools keep changing, so we need to check the current documentation and listen carefully to the clip we've actually received.Google DeepMind: Veo capabilities and limitations

    Knowing when to stop generating

    Another take is always tempting. It helps to decide in advance how many attempts we'll make before stopping to look at the problem again. Perhaps we need a new generation. Perhaps a different cut, a change in sound or a shorter scene would solve it. Repeating the request can become a way of avoiding a decision about the film. More versions won't necessarily make that decision for us.

    Before calling it finished, watch the film straight through at the size and pace intended for the audience. Then make a separate pass for visual glitches, text, sound and export settings. Save the approved cut with its sources and review notes. After all those checks, there is still a simple question worth asking: can we settle into the film, or does something keep pulling us out?

    Sources & further reading

    These are the original sources used in the essay. Tools and their terms change, so check the current documentation before starting work. The suggestions here offer ways to approach a film; they don't guarantee a particular result.