Celebrity Kitchen AI Video: the coffee couple mess, explained

A Celebrity Kitchen AI video is a short vertical clip of two people making coffee together in a modern kitchen — and making a mess of it. One person handles the mug and the coffee powder, the other helps with the kettle and the milk. The powder spills onto the counter, the milk goes everywhere except into the mug, and the two of them react exactly the way the original reel did. The comedy is the fail: two competent-looking people, one simple job, coffee everywhere.

The mechanism is reference-video motion transfer. You do not film anything and you do not describe the scene from scratch: you hand the model a reference video (the original coffee-making reel) and two photos, and it transfers the reference’s exact motion, performance, and camera work onto the two people from the photos [reported]. The kitchen, the actions, and the camera stay; only the faces change.

The wave is four days old: a videosi weekly chart lists it trending since October 27, 2026 [reported], prompt sites vmeditor, vseditor01, and meraprompt published in a cluster on October 29–30 [reported], and two Reels on October 30 use the same coffee-fail reference video, with commenters asking for the prompt [reported]. The trend travels under three names — celebrity kitchen AI video, AI coffee couple video, cofi trend [reported] — and the English head term has not settled yet, so this page covers all three instead of betting on one spelling.

The clip is about 8 seconds [reported], and every viral example follows the same reference reel, because they all transfer the same motion [reported]. Knowing the beats matters: if your generation breaks, it breaks at a specific beat.

The setup — a normal kitchen. Two people at a counter in a modern kitchen, coffee mug, coffee powder, kettle, milk. Nothing unusual yet — the ordinariness is what makes the fail funny.

The pour — coffee into a cup. One person pours coffee into the cup, adds milk and sugar, stirs it around. The other helps with the kettle and the milk. So far, a perfectly ordinary coffee-making scene [reported].

The fail — powder and milk everywhere. The coffee powder spills onto the counter instead of into the mug, and the milk makes a small mess of its own [reported]. The spill is the whole joke: the mess lands on the counter, not on the people, and both of them react — surprise, laughter, a shared look — exactly like the reference [reported].

The reaction — the payoff. The clip ends on the two of them dealing with the mess together. That shared reaction is why the template works for couples: the video is not about coffee, it is about two people being clumsy together.

The template has two parameters — the two people — and that is a variant machine. Any couple, any pair of friends, any two celebrities can be dropped into the same kitchen fail, and each version is personal in the way a generic coffee video never is. The same mechanism drove Tiny Desk (five people) and glowing veins (one letter): one template, infinite subjects.

The second accelerator is that the prompt is copy-paste. The prompt sites all publish the same structure — reference video plus two photos plus one prompt — so reproducing the trend takes zero skill: download the reference, upload two photos, paste, generate. That is why a cluster of prompt pages appeared within two days of the Reels wave — the distribution channel is the prompt itself.

A sibling variant already shows the family: a retro couple-coffee version, framed around a golden-era Bollywood mood — a cozy coffee table, soft light, vintage fashion, slow cinematic camera — runs about 8 seconds and loops cleanly as a Reel [reported]. Same inputs, different mood: the template is a container, and the fail is only the first filling.

You need three files: the reference video of the original coffee-making reel, and one clear photo per person [reported]. Both faces should be clearly visible, front-facing, in good light — the model can only preserve what the photos actually show. Prompt sites themselves note to use only your own photo or photos you have permission to use [reported].

In Flow, start a new project and choose the video workflow. Upload the reference video first, then the two photos, paste the prompt below, set vertical 9:16, and generate [reported]. Prompt sites report a 1–2 minute wait [reported]; when it is done, download it and post it as a Reel or Short.

Create a highly realistic 8-second vertical 9:16 video by transferring the EXACT motion, performance, and camera work from the reference video (@video1) onto the two people in the photos.

Use @image1 for one person and @image2 for the other, keeping each person's face, hairstyle, and clothes from their own photo throughout. Keep the kitchen, the counter, the mug, the kettle, and every action exactly like the reference video.

The two of them make coffee together in a modern kitchen: one pours coffee into the cup, adds milk and sugar, and stirs; the other helps with the kettle and the milk. Then the fail: coffee powder spills onto the counter instead of into the mug, and the milk makes a small mess. Both react with surprise and laughter, exactly like the reference.

One continuous shot. No cuts, no camera jumps, no extra people, no warped hands, no changed faces, no extra objects.

One practical note: the prompt above is our condensed wording of the circulating template, not a copy of any single prompt site’s text. The structure, the beats, and the rules are the same; the phrasing is ours.

The photo slots disagree. Prompt sites map the two photos differently: some assign the first image to the girl and the second to the boy, others assign them to the person on the right and the person on the left [reported]. Before you generate, check which photo lands on which character in your tool — a swapped couple is the most avoidable failure in this template.

Motion drift is the usual failure. The model can improvise instead of copying: prompt-site guidance warns that changing the kitchen background, character positions, camera angle, or original actions produces a result that looks completely different from the reference [reported]. If your clip does not match the reel, your prompt did not lock the reference hard enough — tighten the “exactly like the reference” constraints before blaming the model.

Identity drift, as always. Faces can blend across the two people, and hands, mugs, and kettles can warp mid-motion. The opening photos are your anchor: the clearer and more front-facing they are, the better the model holds both identities apart.

“Free” means their account, not Flow. Prompt sites claim the video is completely free with no subscription [reported], but Flow availability and credits vary by account [unknown]. Treat any “free” claim as that site’s account talking, not a published Flow price.

Celebrity faces are examples, not a license. Viral examples use celebrity photos, but using a real person’s likeness without permission carries platform and consent risk. The safe version of this trend is you and your partner, not two famous faces.

The term is unconfirmed in English. Autocomplete returns nothing for the English phrasings yet, and three names coexist [reported]. This page covers the template, not a bet on one spelling — the same stance as our glowing veins guide.