StudioBuild log

Build log

How We Built a Hotel Lobby AI Generator (and What Didn't Work)

We spent two days trying to turn two selfies into the orange-booth Hotel Lobby duet. This is the honest build log: what we tried, what broke, and the pipeline that finally made both people follow the original moves.

By Hotel Lobby StudioGoogle sign-in · One-time free preview
The quick answer

A Hotel Lobby generator that matches the original performance cannot be one "two photos in, video out" call. In our tests, one-pass reference-to-video models moved the left performer roughly right but left the right performer nearly still. What works: split the original clip into two single-person green-screen performances, generate a first frame of each new person in that exact pose, transfer the motion one person at a time with a motion-control model, then composite both back into the original booth with the original song.

What a Hotel Lobby generator has to get right

The trend clip is two performers sharing a hanging microphone in an orange booth. People recognise it by the gestures as much as by the room: the pointing, the hand roll and the shared "shh" at the end of the line. A generator that keeps the room but invents its own gestures does not read as Hotel Lobby.

  • Two different faces, each staying on its own side for the whole clip.
  • The original booth, microphone and camera push-in.
  • Every gesture of each original performer, at the original timing.
  • Who sings when, matching the original lines.
  • The original song excerpt, in sync with the picture.

What we tried first, and why it failed

  1. One-pass reference-to-video

    We fed MiniMax H3 the two photos plus the original clip and audio as references. A 12-second clip cost us about $0.96 in provider credits. The left performer followed the moves loosely; the right performer barely moved, missed the hand roll at 12 seconds and the "shh" at 15 seconds, and the model sometimes gave the singing to the wrong person.

  2. Rewriting the prompt

    Stricter prompts fixed outfits (keep each person's own clothes) and stopped the model from copying the original performers' build. They did not make the right performer follow the motion. Every render was still a lottery.

  3. Single-person motion apps

    A low-cost character-animation app swapped only one person and returned 832 × 480 video, too soft for a 1080p trend clip.

  4. Premium motion transfer

    A friend's clip made with Higgsfield followed the moves closely, which proved motion transfer was the right idea. It became our quality target.

Why motion transfer has to run one person at a time

Motion-control models animate one character. Kling's Motion Control documentation says that when a reference video contains two or more people, the motion of the person who takes up the largest part of the frame is used. Feed it the full duet and it copies one performer and drops the other.

So the duet has to be taken apart: one clean reference video per performer, one generation per performer, and a final step that puts both back into the booth.

Kling Motion Control API documentation

The pipeline we ship: six steps

  1. Cut the template

    We start from the official 1080p upload and use the 15 seconds from "Let's get it" to just before the wide shot, including the shared "shh". The 10-second option is the first 10 seconds of the same segment. The channel logo is removed.

  2. Isolate each performer on green

    SAM2 tracks each performer through all 383 frames, and a video matting model (RVM) gives soft edges for hair and fingers. Each matte is limited to its own performer, so the neighbour's hand disappears even when they touch. The right performer's crop is widened so his reaching hand is not cut off at the edge.

  3. Build an empty booth

    Both performers are painted out of every frame. Because the camera pushes in, a single still background would show two microphones; a per-frame clean plate keeps the microphone where the camera puts it.

  4. Make and approve the first frame

    Nano Banana Pro places each new person into frame 0 of their performer's pose on green. You check original pose against your first frame, can redo once, and after a redo pick the better attempt per person. Nothing is rendered until you approve.

  5. Transfer the motion

    Kling 3.0 Motion Control animates each first frame with its performer's green-screen reference at 720p. If you add a frontal face photo, it is bound as a Kling face element for a closer likeness.

  6. Composite and add the song

    Each generated performer is keyed off green. Generated green is not flat: a darker halo sits around the person, so the key estimates the backdrop locally and pulls the edge in by 2 pixels to remove the dark rim. Each person is aligned by head position into the original slot. Where the right performer covered the left one in the original, he stays in front. Finally the original song excerpt is added.

What it costs and how long it takes

Motion control is priced per generated second. At Kling's list price of $0.126 per second at 720p, a 10-second duet needs two 10-second passes, about $2.52, before the first frames. Fast or difficult motion can return a shorter clip, and Kling bills the seconds it actually outputs.

Compositing is cheap. In our tests a 10-second duet took about 40 seconds on a 2-vCPU machine and about 80 seconds on a 1-vCPU cloud container. The waits you feel are the first frames and the motion renders.

Kling API video pricing

What still breaks

  • Likeness depends on the photo. A clear, front-lit face with the upper body visible works best; the optional face photo helps.
  • Mouths follow the original performer's mouth movement. It looks like singing the line, but it is not a lip-sync of your own voice.
  • Very fast moves can make the motion model return fewer seconds than requested. We do not deliver a shortened clip; the job fails and your seconds are returned.
  • It is one fixed template. A new trend needs a new cut, new mattes and a new empty booth.

Try the generator

Hotel Lobby Studio runs this pipeline for you: upload two photos, approve the first frames, and download a 10-second ($5.99) or 15-second ($9.99) duet with the song.

Open Hotel Lobby Studio

Photo and prompt guide