How to Make a Consistent One-Minute AI Short From One Phone Photo: Krea's Zoo York Found-Footage Recipe
The hardest thing to fake with AI video is not a monkey on a unicycle. It is the same street, in the same light, at the same distance, twelve shots in a row.
Krea's new guide to Zoo York, a one-minute short about a monkey loose in New York, solves that problem in a way that most AI filmmaking advice skips. It does not chase a sharper model or a longer reference list. It picks a look where the model's weaknesses stop showing: a shaky phone camera filming from across the street at dusk, with the monkey small in the frame and nobody else around.
That choice does more work than any setting. And once you see why, you can borrow it for very different films.
What Krea actually made
The Zoo York guide, published November 5 by the Krea team, starts from one real phone photo of a monkey with a laptop on a New York street. From that single picture, Krea Agent (the chat assistant inside Krea) produced twelve stills, each one a gag: a bursting watermelon, a horse, ducks, an office, a flying carpet, a piano, a rooftop pool, a lowrider, a unicycle, a mattress, a hammock and a bunch of balloons.
Each still became the first frame of a five-second clip. Twelve clips at five seconds each make about a minute. The guide says the whole image set came from five short chat messages.
The tools are named plainly. The stills were made with GPT Image 2.5 at 16:9 and 2K. The clips were made with Seedance 2.5 at 1080p, 16:9, five seconds each.
None of that is exotic. What makes the guide worth your time is the two prompt templates and the reasoning behind them.
The shared line that holds the film together
Every still prompt ends with the same words:
cheap phone photo, taken from across the street, overcast winter dusk, muted colors, monkey small in the frame, empty New York street, no people.
Read that as a cinematographer would. It is a camera and lighting plan in one sentence. "Cheap phone photo" sets the lens and the image quality. "Taken from across the street" sets the distance. "Overcast winter dusk" sets the light, and overcast light is flat, so it hides the small shifts in shadow direction that give AI shots away. "Muted colors" stops one frame from going neon while the next goes grey. "Monkey small in the frame" keeps the face and hands, where drift shows first, too small to inspect. "Empty New York street, no people" removes a crowd of extra faces that would each need to stay the same.
Each phrase removes a decision the model would otherwise make differently every time. That is what consistency is, in the end: fewer decisions left to chance.
The front of each prompt changes only one thing, the gag. Everything that should stay the same rides along at the end, unchanged.
One gag per clip
The video prompt follows a four-line template:
- "Shaky phone footage from across the street, 90s home-video feel."
- The single gag, written as one action. For the opening clip: "The giant watermelon bursts open and splatters juice everywhere."
- "Fast zoom in on the monkey."
- "Audio: real street sounds, no music, no voiceover."
The first line repeats the camera plan for motion. The shaky handheld feel forgives small jitters and warps, because real phone footage has them too. The third line gives every clip the same move, a punch-in zoom, so the cuts feel like one person filming. The audio line keeps the sound under control: street noise only, so you can add music in the edit without fighting a score the model invented.
The second line is where most people go wrong, and the guide says so directly. Ask for more than one action and "the model usually drops part of it." One watermelon bursting works. A watermelon bursting while the monkey runs off and a taxi swerves will lose one of the three.
Cut it like found footage
The edit advice is short and right. Order the clips so the story builds, watermelon first and balloons last. "Use hard cuts. Don't use fades." Then "Add the 90s finish": "Use faded warm color, softer detail, and a little grain."
Hard cuts suit home video, where someone stops and starts recording. The grain and softness pull twelve separately generated clips toward one shared texture, so the small differences between them blend in.
Put this into practice
You do not need a monkey, or Krea, to use this. The method works in any tool that turns a still into a clip. Here is how I would adapt it to your own short.
1. Pick a look that forgives. Before you write a prompt, choose a camera and a world where small errors read as character. Phone footage, security-camera footage, a 16mm home movie, a long lens from far away. Avoid close-ups of faces and hands for your first try.
2. Write your shared line. One sentence that fixes the camera, the distance, the light, the colour and how big the subject sits in the frame. Add the things you want gone ("no people," "no text"). Paste it at the end of every still prompt and never edit it mid-project.
3. List your gags as single actions. One verb each. If a gag needs two things to happen, make it two clips.
4. Make the stills first and judge them as a set. Lay all twelve side by side before you animate anything. If they do not match, change the image model or the shared line, not the individual prompts. Krea did exactly that: its first batch "do not look consistent," and the fix was to ask for "a different image model, maybe GPT 2.5," which the agent used from then on.
5. Animate with one fixed template. Same camera line, same move, same audio line. Only the gag changes.
6. Expect retakes. Budget at least two tries for a few clips. Krea needed "a second try" on two of its twelve, including the hammock swing.
7. Finish in the edit. Hard cuts, one colour grade across everything, a touch of grain, and your own music.
What it costs on Krea
The guide does not say which plan it used, so I checked Krea's pricing page. There is a free plan with 100 units a day. Its plan card says "All video models," but the comparison table marks video access "Limited" for both Free and Basic, and the Seedance 2.5 video counts on the plan cards start at Basic. Paid plans cost $5 (Basic), $21 (Pro) and $63 (Max) a month when billed yearly, or about $9, $35 and $105 billed month to month.
The video estimates on that page do not agree with each other. The Pro plan card says "~65 Seedance 2.5 videos" a month. The comparison table further down says about 23 for a five-second 720p clip on the same plan. The card does not say what clip settings it assumes, and Zoo York's clips were 1080p, which costs more than the table's 720p line. So plan for the lower number. Twelve clips plus a few retakes may use most of a Pro month at full resolution, and a Basic plan will not stretch that far. Watch your unit balance after the first three clips and do the maths from what you actually spent.
Where this recipe breaks
It is a vendor's own story. Krea wrote the guide to show off Krea. It does not publish every still prompt in full, the agent's first image model is not named, and nobody outside Krea has reported repeating the result.
The look is the trick. This works because the subject is small, the light is flat and the camera is cheap. Try the same method on a close-up dialogue scene and the drift comes back. If your film needs faces, you need a different approach, such as locking a character with reference stills.
No dialogue, no story beats inside a clip. Five seconds, one action, street noise. Anything longer or talkier needs more structure than this.
Rights still matter. Start from a photo you took or have permission to use. The guide's own photo was real, and an AI short built on someone else's picture inherits their claim to it.
The meter runs on retakes. One gag per clip keeps retakes down, but at 1080p every retry costs real units, and the pricing page cannot tell you exactly how many.
The idea worth keeping
The easy lesson from Zoo York is "use these prompts." The better one is that consistency in AI video starts with an art direction choice, not a model choice. Pick a world where the camera is far away, the light is soft and the footage is supposed to look rough, and the model's habit of changing small details stops being a flaw.
So before your next short, write the shared line first. Decide what the camera is, where it stands and what the light is doing. Then try it on three gags, not twelve, and see whether the set holds together. If it does, you have a film. If it does not, change the sentence, not the shots.
- Sources: Krea: Zoo York guide · Krea pricing