Gangnam Style AI video prompts that actually hold a face

Most prompts for this trend fail for the same reason: they describe the song instead of the shot.

Write "Gangnam Style music video" and the model reaches for the elevator, the horse statue, the blue suit. You get a parody of a music video instead of a phone clip of three people dancing badly on a Tuesday night.

Describe bodies, light and camera. The meme reads on its own.

The five blocks

Every prompt on this site is assembled from the same parts:

1. Framing — "Vertical 9:16 handheld phone footage". Locks aspect ratio and the phone-shot feel.
2. Reference — "The person from @Image1". Anchors identity instead of describing a face in words, which never works.
3. Scene — concrete nouns: neon shop signs, wet asphalt, passing headlights. One sentence.
4. Motion — name the body parts. Hand on imaginary reins, hips shifting, shoulders rolling, weight moving foot to foot, on the beat.
5. Exclusions — no text, no logos, no extra people in frame, natural skin texture.

Order matters less than completeness. A prompt missing block 5 will invent a crowd; a prompt missing block 4 will produce three people standing still.

The default prompt

Vertical 9:16 handheld phone footage at night on a busy city street.
Three friends dance side by side in the middle of the road, performing
the same loose horse-riding-style bounce on the beat: one hand on the
imaginary reins, hips shifting, shoulders rolling. The person in the
centre is the reference from @Image1. Neon shop signs and passing
headlights glow behind them, wet asphalt reflects the light. Slight
handheld shake, occasional push-in, natural skin texture,
photorealistic, 24fps motion blur. No text, no logos, no watermark.

Variants worth keeping

Rooftop. Swap the street for a skyline of lit windows, cool ambient light with warm practicals behind. Wider framing. Works better with two people than three.

Wet alley. Magenta and cyan neon reflecting on the pavement, light rain in the beams, chest-height camera, slow push-in. More cinematic, fewer extras.

Solo. Drop the friends, add "no extra people in frame". This is the easiest version to keep consistent, because one reference anchors one identity.

Pet or cartoon. Replace the person with "the character from @Image1", exaggerate the bounce, keep the neon street. Comedy scales better than realism here.

Deadpan office worker. Business suit, neutral expression, perfect timing. The joke is the contrast, so keep the face still and let the body do the work.

The negative prompt

extra fingers, deformed hands, duplicate faces, warped text, watermark,
logo, subtitles, cartoon shading on a photoreal render, oversaturated
skin, plastic skin, crowd of strangers in the foreground, camera cuts,
split screen

Not every model accepts a separate negative field. If yours does not, append the list to the end of the prompt after "Avoid:".

Which model to use

• Google Flow / Gemini (Veo) follows detailed prompts most literally. Clips are short, so stitch two for a longer take.
• Kling 2.x produces the most convincing dance motion. Keep motion strength low-to-mid or the limbs get theatrical.
• Seedance 2.5 holds a face longest in a single take, which matters if you want one unbroken nineteen seconds.
• Higgsfield is preset-driven — fastest if you do not want to write anything.
• CapCut / Dreamina templates animate a still. They cannot generate a dancing body, no matter what the prompt says.

Three things that waste credits

Re-rolling instead of fixing the photo. If the face drifts every time, the source is the problem.

Asking for three faces with one reference. One photo anchors one identity. Add @Image2 and @Image3 or accept two invented strangers.

Chasing length. Drift compounds. Get eight seconds right before you ask for nineteen.

What to do with the output

Export silent. Upload. Attach the track from the platform's licensed sound library. That is the only version that stays monetizable — and the only version that does not get you a strike.