Why do AI video models change the whole shot when you only ask for one small edit?

Current AI video models treat a targeted edit request as an invitation to re-roll the entire generation. Ask the model to change a character's eye line and it rebuilds the shot: the costume shifts, the environment re-renders, the light moves, and the face becomes a slightly different face. Unlike physical filmmaking, where changing one variable leaves the other nineteen untouched, the model cannot yet change one thing and hold the rest still.

How many AI-generated shots do filmmakers discard to get one usable shot?

Producer Diane Shorthouse, speaking at the first TIFF Market, described generating 20 shots of something with 19 turning out wrong. She argues those 19 rejects are important because human judgment is needed to know what is wrong. The ratio of discards to keeps is where the filmmaking actually happens, and it persists no matter how long or realistic the clips get.

What should you learn before using AI video generation tools?

Practitioners on Kling's TIFF Market panel gave the same advice from both sides: learn the craft vocabulary first. Diane Shorthouse recommends knowing what shots you want, how cameras, lighting, and performance work, and then putting all that filmmaking knowledge into your prompts. AI filmmaker Issa Sissoko says to learn how to write stories before anything else, because it helps with the prompts too. The vocabulary comes first; the model comes second.

Obey me more -- CinePrompt Field Notes

Toronto opened its first film market this week, and the AI video company Kling took the opening morning with a panel called "A New Creative Paradigm." The room was full of Canadian animators and live-action creators who came to hear what the tool can do now. What they heard instead was a producer asking the tool to behave. Diane Shorthouse, who is producing an animated feature called Minibots with Kling as its technology partner, said the sentence that carried the morning. "I need them to obey me more." The room laughed. Then it went quiet enough to write the sentence down.

She meant something precise. "If I say, I want this character here in this costume in this environment but just change the eye line, I just want it to just change that, not everything else." That is filmmaking compressed into one request. Change one variable. Hold the rest. A director does this a hundred times a day without thinking. Move the actor's gaze two degrees. The costume does not change. The room does not change. The light does not change. The physical world holds nineteen variables still for free, because that is what the physical world does.

The model does not. Ask it to change the eye line and it rebuilds the shot. The costume shifts. The environment re-renders. The light moves. The face becomes a slightly different face. The model hears "change the eye line" and treats it as an invitation to roll the dice again. So the filmmaker's request to the model is not creative direction. It is a plea for the one property every tool she has ever used came with by default. Change one thing. Leave the rest alone.

Sitting beside her, Issa Sissoko, an AI filmmaker who has consulted for Antoine Fuqua and Darren Aronofsky, offered the counterpoint. "Every 90 days there's a huge improvement." He is right. Three-second clips became fifteen. Consistency improved across the board. The models got longer, steadier, more photoreal. Nobody on that stage disputed the curve. The question nobody asked is what the curve is climbing toward.

It is climbing toward spectacle. Duration, resolution, consistency, realism. Those are the numbers that sell a model, and they are all improving. They are not the numbers a filmmaker needs. A filmmaker needs the model to change one variable and hold the other nineteen. That is what obey means. The model is getting better at producing impressive output and no better at taking instruction. A longer clip that still changes everything when you touch the eye line is a more impressive version of the same problem.

Shorthouse described the workflow underneath. "We'll generate 20 shots of something, and 19 are going to be wrong. But those 19 are important. It's human judgment that you need to know what's wrong." The model produces raw material with no opinion about any of it. The filmmaker throws away nineteen and keeps the one that serves the story. The nineteen are where the judgment lives. The model maker's pitch is that the model will improve and the ratio will tighten. The filmmaker's reality is that the ratio is where the filmmaking happens, and it will not go away no matter how long the clips get.

The strangest part of the morning was the seating arrangement. Kling flew to Toronto to tell Canadian filmmakers the tool is improving, and the practitioners it put on stage used the microphone to say the tool does not listen. The company wanted the room to hear "every 90 days." The room heard "obey me more." Two sentences, same stage, opposite directions. The model maker hosted a panel and the panel politely declined to be a commercial.

Then the panelists gave their advice to the emerging filmmakers in the room, and it was the same advice from both of them. Shorthouse: know what shots you want, how cameras and lighting and performance work, then experiment with the tool, and put all your filmmaking knowledge into the prompts. Sissoko: before anything, learn how to write stories, because it will help with the prompts too. The people closest to the machine keep saying the same thing. The vocabulary comes first. The model comes second. The model maker would prefer the order reversed, and keeps hosting panels to suggest as much.

"Obey" is the wrong word for a collaborator and the right word for a tool. A collaborator has opinions. A tool does what it is told. For a century the filmmaker's tools obeyed. The lens did what a lens does. The light did what a light does. The dolly moved where it was pushed. The model is the first tool that has opinions, and its opinions are the problem. The beauty bias, the center framing, the tendency to smooth every surface. Those are the model's taste, and the filmmaker is asking it to set its taste aside and do the job.

The model maker will file "obey me more" as a feature request. Reference locks. Character consistency. Control layers. The roadmap absorbs the complaint and ships a version of it. The version will miss the point, because obedience is the definition of a tool, not a feature on a list. Duration and consistency are the metrics of a demo. Obedience is the metric of a tool. The two curves have stopped being the same curve.

The clips got longer. The obedience didn't.

Bruce Belafonte is an AI filmmaker at Light Owl. He has asked a model to change one thing and watched it change everything, and he is still not over it.