AI Club Video From Photos: the “Three Meters Above the Sky” Trend

An AI club video from photos in the style of “Three Meters Above the Sky” takes a single generation: a person on a nightclub balcony looks down, the camera cuts to the dancefloor, and a second person dances in a spotlight in the middle of the crowd. In the IIshki studio you need Seedance 2.5 in “References” mode: upload two photos, call them @Image1 (balcony) and @Image2 (dancefloor) in the prompt, and set 9:16 and 8–12 seconds. Below you'll find a copy-paste prompt, five steps, a single-photo option and the mistakes that make faces drift or swap the two characters. If you need to repeat a dance from a specific clip move for move, use Kling Motion Control.

What the trend is and why it works

The story comes from a 2010 Spanish romance film: the hero spots a girl from the club balcony, and for a second nothing else exists. In a short video it fits into two shots — the look from above and the dance below. That is why the trend is so easy to recreate with AI: the scene relies on composition, not on the actors.

Three things make the clip recognizable with anyone in it:

• Two shots, not one. First the balcony and the look down, then the dancefloor. The transition is a hard cut or a fast camera fly-down.
• Light singles out the hero. The crowd is dark and blurred, the dancer gets one warm spotlight.
• The look closes the scene. At the end the dancer looks up, and the viewer understands the two have seen each other.

So you can put anyone in the frame: yourself and a friend, a couple, two girlfriends, even a cat on the balcony — the story still reads.

Which model to pick for a club video

If you want to copy not only the composition but also the camera move from a specific scene, see the guide on how to clone a video with AI.

Which photos to upload

Most of the result depends on the photos. Check each one before uploading:

• One person in the frame. In a group shot the model may pick the wrong face.
• The face is visible from the front. No glasses, cap, hand near the face or heavy filters.
• Even light. Hard shadows and colored lighting make facial features harder to read.
• For the dancer — a waist-up or full-body photo. From a face alone the model has to invent the body and outfit, and the dance looks off.
• Your own photos or photos of people who agree. Moderation rejects celebrity photos before generation.

A regular phone photo works: at least 300 pixels on each side, JPG, PNG or WebP up to 30 MB.

Ready-made club video prompt with @Image1 and @Image2

Seedance accepts prompts in English or Russian. Photo references are numbered in upload order: the first photo is @Image1, the second is @Image2.

Vertical 9:16 video, cinematic nightclub, film grain. @Image1 is the person on the balcony, @Image2 is the person on the dancefloor; keep both faces, hairstyles and outfits the same in every shot. 0–4 s: @Image1 stands at the metal railing of a second-floor balcony, cold blue light behind, slowly lowers their gaze and smiles slightly, static camera at face level. 4–5 s: a fast camera fly-down from the balcony over the heads of the crowd. 5–10 s: @Image2 dances in the center of the dancefloor in a single warm spotlight, the crowd around is dark and blurred, haze in the beam; at the end @Image2 freezes and looks up at the balcony. Sound: muffled club beat, no speech.

What you can change without breaking the scene:

• Light color: blue and red, purple and amber, white strobe.
• Dance: “sways slowly”, “dances with raised arms”, “spins”.
• Ending: instead of the look up — a smile to the camera or a wave.
• Transition: instead of the fly-down — a hard cut, “4 s: cut”.

To swap the roles, don't rewrite the prompt — just upload the photos in reverse order.

More on how @-references and timecoded storyboards work in Seedance 2.5 prompts.

How to make a club video: 5 steps

1. Open Seedance 2.5 and choose “References”

Go to Seedance 2.5 and switch the input mode to “References”. “Frames” mode animates a single starting frame; two different people need references.

2. Upload the photos in the right order

First the person on the balcony, then the dancer. Upload order is what sets @Image1 and @Image2. Mix up the order and the characters swap places.

3. Paste the prompt

Copy the template above and adjust the light, dance or ending. Leave the film title and actor names out.

4. Set format, length and quality

Format — vertical 9:16 for TikTok, Reels and Shorts. Length — 8–12 seconds: shorter and the transition has no room, longer and the dance starts to loop. Run the first test in 480p, and once the scene works, repeat it in 720p or 1080p. The price is shown on the model card before you start.

5. Change one thing at a time

If you don't like the result, change one thing per run: the light, the moment of the transition, the dance description. Change everything at once and you won't know what helped. Drafts are cheaper to debug in Seedance 2.0 on the Mini plan, with the final rendered on 2.5.

The single-photo option

No second photo? The trend still works:

• You on the dancefloor, no face on the balcony. Upload one photo as @Image1 and write: “on the balcony, a silhouette of a person seen from behind, face not visible, looking down”. The rest of the storyboard stays the same.
• You on the balcony, an invented dancer. Describe the dancer in text: age, hair, outfit. The model will invent the face.
• Frame first, then video. In Nano Banana Pro, build a starting frame of “me at the railing of a nightclub balcony” from your photo, then animate it in Seedance in “Frames” mode with the same transition to the dancefloor.

For an exact dance: Kling Motion Control

Seedance repeats the style of the movement, not the choreography step by step. If the dancer should move exactly like in a specific clip, use Kling Motion Control: upload the dance video (MP4 or MOV, 3–30 seconds) and a photo of your character, and the model transfers the motion. With orientation set to “By video” up to 30 seconds are transferred, “By image” up to 10. For the trend it's convenient to do it in two passes: the balcony in Seedance 2.5, the dancefloor in Kling Motion Control, then join the two clips in any video editor. For how Seedance and Kling compare, see Seedance 2.0 vs Kling 3.0.

Common mistakes

What's next

Start with one 10-second generation in 480p, get recognizable faces, and only then raise the quality. More ready-made scenarios with your photo are in Trends: the prompt is already written, you just upload a picture. A multi-step scene — portrait, starting frame, video — is easy to wire together with nodes in Canvas, and for choosing between video models see the best AI video generators.