On September 6, a 54-year-old former nonprofit worker from Canada collected the top prize at the Reply AI Film Festival on the Venice Lido. Eight days later, Kling hosted a panel at the Toronto International Film Festival's new market where a producer with more than 30 feature credits explained her process. Between the two, you get a clear answer to the question we hear most: how filmmakers actually use AI video when real money or a real festival is on the line.
The short answer is that they treat the model like a camera department that never stops arguing. They write the script first, break it into shots, lock the cast with reference sheets, and then generate far more takes than they keep. Nobody on either stage described typing a paragraph and waiting for a film to come out.
How filmmakers actually use AI video on a festival short
Robert Gaudette's "A Face Only A Mother Could Love" runs 7 minutes 50 seconds and follows Marcel, a solitary Parisian dancer with a facial disfigurement. It beat more than 3,000 submissions from 76 countries, according to the festival's announcement, and it had already taken the Grand Prix at Runway's AIFF earlier this year. The Reply release credits Grok, Nano Banana and Seedance. The Hollywood Reporter's profile from the Runway win puts Runway in the mix as well. He used whatever got the shot.
The method matters more than the tool list. Gaudette wrote a detailed prompt for every shot, specifying character design, camera movement, palette and emotional register. He generated character sheets before any footage, so Marcel's face and his particular way of moving stayed coherent across the film. And he threw away hundreds of generations to protect one thing: the character's gentleness. His line: prompting is not directing. Directing is the curation.
The producer's version
Diane Shorthouse, co-founder of Evolutionary Films and producer of the Kling-backed animated feature "MINIBOTS", was blunter on the Toronto stage. "The script is still the blueprint, so I'm not thinking about prompts, I'm thinking about shots," she said, per Variety's coverage of the panel. Her checklist is the one every producer already carries: where the camera goes, what the performance needs to be, what it cuts to, and which assets are required to get there.
A producer with 25 years of credits is describing a shot list and an asset list, then handing both to a model. The prompt is the last step, and the least interesting one. "MINIBOTS", announced at the Cannes market in May with Kling as technology partner, also runs under what Variety described as a performance-first charter: every performance stays human-created and actor-owned.
Previs is the quiet big use
Issa Sissoko, founder of Amissa Studio, sat on the same panel. He has led previs and hybrid-production consulting on projects with Antoine Fuqua, Doug Liman and Darren Aronofsky, plus commercial work for Coca-Cola, Netflix and Amazon.
His description of hybrid production is the most useful thing said at either event. The hard problem is taking a real actor's performance and changing the environment around it without touching the performance. Getting the model to keep the performance intact, let alone add to it, is "very tricky" in his words. So the goal today is to reduce how much of the final frame depends on what the model invents.
The other job is previs. Sissoko uses AI video to let directors and producers see a sequence before anyone rents a crane, which surfaces problems early and tells them where the real production budget should go. That is not a demo reel use case. It is a line item on productions that will never be marketed as AI films.
A workflow you can copy this week
What all three described, compressed to a single scene:
- Write the scene as script, not as a prompt. Dialogue, action, one line on tone.
- Break it into shots with a real shot list: size, angle, movement, what it cuts to.
- Generate a character sheet per character (front, three-quarter, profile, one expression) and attach it to every shot.
- Generate 4 to 6 takes per shot. Keep one. Log why the others failed.
- Cut the keepers before generating anything else. Only then go back for pickups.
Step 4 is where most first-time creators quit, because it feels like the tool is failing. It is the normal ratio. Gaudette's hundreds of rejected generations are the same ratio at festival scale.
We built Promvie around that order of operations. The script comes first, the cast is locked as reference sheets before any shot is rendered, and every scene is generated against a shot list rather than a single prompt. The how it works page walks through it, and our earlier post on how AI filmmakers keep characters consistent covers the reference-sheet step in detail.
What to do about it
If you are making a short for a festival deadline, borrow Gaudette's discipline: character sheets first, one prompt per shot, and a willingness to bin most of what comes back. If you are pitching a client or a studio, borrow Sissoko's: sell the previs, keep the actor's performance untouched, and let the model fill in the world.
The people winning with these tools right now are not better prompters. They are better at the part of filmmaking that existed before the prompt box did.