How to make the Gangnam Style AI video

The trend looks effortless on the feed: three friends, a neon street, ten seconds of the horse-riding bounce. Then you try it and get melted hands, a stranger's face and a background that keeps re-arranging itself.

Almost none of that is bad luck. It comes from three decisions made before you press generate.

1. The photo decides the ceiling

An image-to-video model has exactly one frame of information about you. Everything else — the back of your head, your body proportions, how your clothes hang — is invented.

That is why a tight, passport-style face crop performs so badly. The model has no shoulders to attach arms to, so it improvises, and improvised arms are where the uncanny valley starts.

What actually works:

• Front-facing, eyes visible. Three-quarter angles force the model to hallucinate the hidden half of your face, and it shows.
• Head, shoulders and upper body in frame. The crop is the single biggest quality lever.
• Even light. Window light or overcast daylight. Hard side shadows read as two different faces.
• Sharp. A compressed screenshot of a screenshot loses the micro-detail the identity pass depends on.
• One person. Group photos make the model pick a face, and it often picks wrong.

Phone HEIC files are fine — most tools convert them in the browser before upload.

2. Pick a scene, not a mood board

The format only has a handful of visual signatures: neon signage, wet asphalt, a handheld camera at chest height, and light coming from behind the subject rather than on them.

If you are prompting manually, say those things explicitly. If you are using a generator, pick the closest preset and stop fiddling:

| Scene | Reads as | Best for | | --- | --- | --- | | Night street | Authentic phone clip | The default. Start here. | | Rooftop | Music-video cut | Two people, wider frame | | Wet alley | Cinematic | Moodier posts, fewer extras |

Daylight versions exist, but they lose the contrast that makes the format legible. A sunny street dance looks like a stock advert.

3. Test at ten seconds

Identity drift compounds. A face that is 95% right at second three can be visibly wrong by second fifteen, because each frame conditions on the previous one.

So render the short version first. It costs half as much and exposes every failure mode you would have hit later: hands, extra people, wardrobe changes, the slow slide into a different person.

Only spend the full nineteen seconds once the short pass survives a pause at the halfway mark.

The five failures, and what to do

It looks like my cousin. Angle or resolution. Straight-on, sharper, whole head visible. Do not re-roll five times on a bad source photo — five re-rolls cost more than one better selfie.

Hands melt or multiply. Standard for current models at this length. Keep hands out of the source frame, prefer wider shots where hands are small, and re-roll rather than trying to prompt the problem away.

Extra people appear. Crowd density is the biggest source of flicker. Add "no extra people in frame" and avoid busy street photos as reference.

The dance is stiff. Motion needs cues. Name the body parts: hand on imaginary reins, hips shifting, shoulders rolling, weight moving foot to foot, on the beat. Presets already carry this language.

It failed and I was charged. A failed or safety-blocked job should return credits automatically. If it has not after a few minutes, contact support with the task ID — that is what it is for.

Posting without a strike

Download the silent MP4. Upload it. Then add the track from the platform's licensed sound library — TikTok, Reels and Shorts all have it.

Do not download the song as an audio file and mux it in. That is the version that gets muted, and it is the reason every reputable tool in this niche ships silent renders.

Keep the caption descriptive and honest: three friends, one night street, AI. Do not imply it is the official video, and do not tag the artist's account as a collaborator.

The short version

Front-facing photo with shoulders. Night street. Ten seconds first. Silent export. Sound added in-app.

That is the whole method. Everything else is taste.