The onion rings came out of the fryer at 10:40 on a Tuesday and I had until 10:47 before the lunch rush made the pass unusable. One phone, no light kit, a manager watching the clock. That is the actual working condition for most food video, and it is why so much of it looks like an apology.

I shoot short clips for small restaurants — one-location places, a couple of trucks, one bakery that still writes its specials on a chalkboard. Over the last stretch I have moved most of that work into an AI video maker instead of trying to stage motion in a kitchen that is about to get busy. The method below is what survived. It is not a creative philosophy. It is a sequence.

Start From the Photo You Already Have

Most of what an AI video maker gets wrong traces back to a weak first input. Every place I work with has a decent still of the thing they sell. Usually it is on the wall menu or the delivery app.

Use it. A generated clip that starts from a real photograph of the real item keeps the sesame seeds where the sesame seeds are. Start from a text description instead and you get a plausible burger that is not the burger — a bun with the wrong shine, three pickles where the kitchen puts two. Customers who order that item every week notice, and they are the ones you cannot afford to confuse.

So the first input is a photograph, not a sentence. The sentence comes second, and its job is only to say what moves.

Give the Model Two Frames When the Shot Has to Travel

Most food motion is small. Steam rising. Cheese pulling. A hand setting a basket down.

Movement like that needs one frame to pin it down and a short instruction to describe it. But some shots need to arrive somewhere specific — closed box to open box, whole pie to cut slice. For those I use first-and-last-frame mode, which takes a starting image and an ending image and generates the travel between them. Two photographs you can shoot in ninety seconds, and the result lands on a state you chose rather than one the model invented.

That model also accepts as many as five reference images and three audio files in a single generation, with each asset given a stated role: this one is the product, this one is the room, this one is the voice. Assigning roles matters more than adding assets. Five references pointed at five different jobs behave. Five references all shouting “make it look like this” fight each other.

Fix the Length to the Action, Not the Platform

No AI video maker will warn you that a clip has run past its point. Every platform will happily take the longer file. Neither is a reason to make one.

Steam rising is a three-second event. A hand-off is four. If the action finishes and the clip keeps running, the viewer watches a still photograph that is technically a video, and the drop-off shows up in the numbers the following week. Whole-second control helps here: set the duration to the length of the thing that happens, then stop.

The exception is a sequence with real steps — a build shot, a walkthrough of a prep line, an order going from window to bag. Those earn room. Thirty seconds in a single pass is available when the story genuinely has that many beats, and it can be continued across further rounds without the hands changing shape or the counter changing colour halfway through. Fifty reference assets can be joined into one generation on that model, which sounds excessive until you are matching a wardrobe, a logo placement and a room tone at the same time.

Most weeks I do not need it. When I do, nothing else substitutes.

Plan the Sound Before You Generate, Not After

I lost half a day once because I treated audio as post. The clips were fine. The audio I dropped underneath was a stock kitchen loop with a different room in it, and every clip suddenly felt dubbed.

An AI video maker that produces its own audio changes the order of operations here. Decide the sound at the same moment you decide the picture. Name what you want — the fryer, the register, one line of dialogue, a room that sounds like the room — and check duration, aspect ratio and sound as one decision before submitting. Reviewing them separately is how you end up with a correct clip in the wrong shape.

On aspect ratio, one practical note: pick the primary placement first. A vertical crop of a horizontal plate shot cuts the plate. Generate for where it will actually run, then make a second version deliberately if you need one.

What This Does Not Replace

A person is still a person. If the owner is on camera, film the owner. A face that lands close but not exact will cost you more trust than shaky handheld footage ever would, and regulars feel that wrongness long before they can name it.

Same for anything a customer will hold up against what arrives on the tray. Generated food that is prettier than the real plate is a complaint, not a campaign. I keep generated work on atmosphere, motion and B-roll, and I keep the hero item honest.

The rule I give every operator: shoot one real photograph of the item today, before you open any AI video maker. Everything downstream gets easier, and the clip that comes out will still be recognisably yours.