Hotel Lobby AI Video Results: Our Tests, Fixes and Limits

Until October 3, the booth was re-shot from a clip of the original Hotel Lobby performance. On Wan 3.0, that clip pulled the original performers into our tests: the French bulldog turned into one of them at a camera cut, and both babies were replaced. The babies were replaced on Seedance models too.

What we changed on October 3: a new reference clip performed by two AI-made stand-ins, a redraw of each photo into a standing studio portrait, and a rewritten prompt. Re-run with the same three pairs, all three kept their faces, one try each.

• Rings from the original performers appeared on a baby’s hands.
• The stand-ins in the new clip wear black gloves, and the first runs copied them onto a man’s hands and a dog’s paws. A line asking for no gloves removed them.
• The old clip cut every 3 to 4 seconds, and every video copied the cuts. Swapped faces started exactly at a cut.

• A dog photographed sitting stayed seated in an early test. The portrait redraw now stands every performer up.
• A tongue-out photo kept the tongue in the portrait until we asked the redraw for a closed mouth.
• Held items and small jewelry drop out in the redraw.

The photo checklist built from these findings is in which photos work best.

The reference clip is wide, and it moves in close twice. In 9:16, the close shots cut off the person on the right. On October 4 we added a vertical version of the reference clip for 9:16, 3:4 and 1:1. Three 15-second tests at 480p (two babies, the man and the bulldog, the bride and groom) kept everyone in frame, except for about one second near the 10-second mark when the right performer brushes the edge.

Seedance AI raps that carried the original track in the reference clip were blocked by the video provider for audio copyright, 3 times out of 3. With the clip silent and the track sent only as a rhythm reference, 4 out of 4 went through. Since October 5, the AI rap on the Hotel Lobby stage writes a new beat in the track’s rhythm, with new words.

In beat tests on October 4, one of six AI raps came back with almost no music. The beat description was rewritten on October 5.

• After the fixes, each test case passed on one try. The same inputs give a slightly different video every time.
• About one second of edge clipping in vertical videos.
• Babies look a little older once stood up; dog paws turn hand-shaped; sizes are not to scale.
• In one of our AI raps the camera pulled back and showed the studio floor for a few seconds, although the prompt asks for a fixed camera.
• Not tested yet: one photo of two people split by the redraw, 720p and 1080p in the new setup, 5 and 8 second lengths, cats, two pets, voice notes, uploaded clips and raps in other languages.

Every new try is paid like the first: 6 credits for a 12-second video at 480p. A render that fails is refunded automatically; a finished video you do not like is not. Current prices are on the pricing page.