How to Create a Two-Person "Hotel Lobby AI" Rap Video from Photos Without Face Drift

Two-person music and rap templates are everywhere right now. The one that keeps showing up is the "Hotel Lobby AI" format: two people in a matte-orange studio booth, hanging mics, a short punchy duo performance over a recognizable beat. It looks simple, but it is one of the easier formats to ruin with a generic image-to-video model.

This post covers why it breaks, how to prepare the input photos so it doesn't, and one workflow that avoids hand-writing prompts.

Why two-person videos drift

Most image-to-video models are tuned for a single subject. Put two faces in one frame and you tend to get:

• Identity bleed: features from face A leak into face B. Hair, jawline and skin tone start to average out over a few seconds.
• Morphing: a face quietly changes shape when the head turns or the mouth opens wide, which is exactly what rap performance does all the time.
• Lip-sync mismatch: the mouth movement follows the beat for one person and lags for the other, because the model treats the pair as one blob of motion.

A text-to-video prompt makes this worse. With no reference photo, the model invents two faces, and nothing anchors them to real people.

Prepare the input photos (4 rules)

Most failures come from the inputs, not the model. These four rules fix most of them:

1. Two separate single-person portraits. Do not use one group photo. Each face gets its own clean reference.
2. Front-facing, eye-level. A roughly straight-on angle gives the model the most reliable geometry. Big side profiles and extreme high or low angles are where morphing starts.
3. Even lighting. Avoid harsh side shadows and strong backlight. If one photo is warm indoor light and the other is cold daylight, expect the skin tones to drift toward each other.
4. Nothing covering the lower face. No sunglasses, no hand on the jawline, no mask. The mouth and jaw have to move believably, so the model needs to see them.

A quick check before you generate: crop each photo to the head and shoulders. If you can't clearly see both eyes and the full mouth, pick another photo.

A prompt-free workflow

Instead of writing a long prompt describing the booth, the mics and the camera move, you can use a template that already encodes the scene and the timing, then only supply the two faces.

I've been building this into ClipMe AI: the Hotel Lobby AI Duo Video Generator takes two portraits and produces an approximately 15-second duo performance synced to the original beat. The Portrait Preparation & Tutorial page walks through the photo rules above with examples.

Two design decisions worth explaining, since they affect whether you'd want to use something like this:

• One-time credit packs instead of a subscription. A short, format-driven video is something most people make a handful of times, not every day. Credit packs start at $9.90, credits don't expire, and a failed generation is refunded automatically, so there is no monthly charge to remember to cancel.
• Private storage with automatic deletion. Uploaded portraits are faces, so they go to a private bucket (not public URLs), photos are deleted after 7 days, videos after 14 days, and nothing is listed in a public gallery.

Takeaways

• Two-person templates fail mostly because of identity bleed, and the fix starts with the input photos.
• Use two separate, front-facing, evenly lit portraits with the lower face visible.
• If you don't want to write prompts, use a template that fixes the scene and only takes the faces.

If you've found other input tricks that reduce face drift in multi-person video, I'd like to hear them in the comments.