The Hotel Lobby AI prompt, explained
Updated 15 October 2026
You never type a prompt into this product. The instruction is composed for you, because the things that make this effect work are the things a casual prompt leaves out. This page explains the shape of that instruction, so you can judge a clip — or build your own version in another tool.
Three jobs, three blocks
The instruction does three separate jobs, and mixing them up is the usual reason a render fails. Each input has to be told what it is *for*:
- The identity block names the two photos and says what they supply: the face, the hair, the clothing. It also fixes which photo is the performer on the left and which is on the right, and forbids the two from blending.
- The performance block names the reference video and says it supplies the choreography, the camera and the timing — and explicitly that the people already in that video must not appear in the result.
- The room block names the fixed elements: the orange walls, the single microphone hanging from above on its cable, the lighting. Anything not named here gets invented.
The rule that costs the most: say it out loud
A video model does not treat an unmentioned element as "leave it as it is". It treats it as "make something up". An early version of our own instruction never mentioned the microphone, and the model confidently produced a microphone on a stand — wrong scene, wrong joke.
The same rule applies to the photos. If the instruction does not say that the uploaded images are the cast, the model will happily keep the performers it already has and ignore the uploads entirely. Two runs of the same inputs can differ purely because one prompt stated the casting and the other assumed it.
Wording is a control signal, not just description
Some words change what kind of job the model thinks it is running. Instructions phrased as "edit this video", "replace the person" or "add a second singer" can be routed as video-editing work with different constraints, which fails in ways that look like a quality problem but are really a classification problem.
The way around it is to describe what the clip *is* rather than what to do to the footage — a performance in a room, with two named performers — and to keep "replace", "delete" and "add" out of the instruction entirely.
A prompt you can adapt elsewhere
If you are building this in another tool, the structure below is the one that has worked for us. Name your own inputs in place of the bracketed parts, and keep the three jobs separate.
- "The reference video is the performance to follow, and the two reference images are the cast."
- "Image 1 is the performer on the LEFT; image 2 is the performer on the RIGHT. They keep their own faces, hair and clothing, never swap sides and never blend."
- "The people already in the reference video must not appear anywhere in the clip."
- "Follow the reference for the choreography, the body movement, the camera framing and the timing. Keep every gesture at its original time and speed."
- "Keep the orange booth, the single microphone hanging from above on its cable, and the lighting. Do not copy the background of the reference photos."
What no prompt can fix
Prompting cannot rescue a photo where the face is a few pixels wide, and it cannot make a model honour a beat it has decided to ignore. That second case is why the honest answer to "why is it not frame-perfect?" is: this class of model follows a performance, it does not clone it. Expect the gestures and the timing to track the original; expect the details to be its own.
Questions
Does this product let me write my own prompt?
No, and that is a product decision rather than a missing feature. Free-form prompts are a different tool for a different person — they trade reliability for control. Here the instruction is fixed so that the result is recognisably the same trend every time.
Why publish the prompt at all?
Because the useful part of this category is the method, not the secret. Anyone can spend a few dollars learning it; being straight about how it works is a better reason for you to use this version.
Make one with your own two photos
Uploading and previewing are free — you only spend credits when you press Generate, and a clip whose faces cannot be verified is refunded automatically.
Face can’t be verified? Credits are refunded.