How to Make a Rap Duo Video With AI, Start to Finish

What you need before you start

You are assembling three things: two faces, one mood, and one reason. Everything else is a choice the model can make better than a longer prompt can. Get those three in place and the render handles the staging, the performance and the sound.

Resist the urge to write a script: the model writes the vocal from your cue, and an 8- to 10-second clip can only carry so much. A name plus one detail beats a paragraph.

• Two front-facing photos, one per performer, taken in even light.
• A stage: Hotel Lobby, Luxury Lobby, Studio Booth or Street Cypher.
• A short cue: both names plus the occasion in one line.
• Permission from everyone in the photos — including yourself.

Step 1: Pick two photos that survive the render

Each upload slot expects exactly one face. A portrait with the face filling a decent part of the frame maps more cleanly than a group photo or a distant shot. If your best photo is a crop, crop it closer before uploading.

Light matters more than camera. Flat daylight or a bright room beats a dramatic backlit silhouette, and a recent photo beats one from five years ago. Skip sunglasses, heavy beauty filters and side profiles: all three hide the mouth or the eye line that the render needs.

• Face the camera with both eyes visible.
• Even light — window light or a bright room, not one hard side lamp.
• Recent photos, ideally from the last year or two.
• One face per slot. No group shots, no screenshots.

Step 2: Choose the stage before you write the cue

Stage choice does most of the art direction. Hotel Lobby gives you the orange late-night TV-studio look. Luxury Lobby gives you marble, brass and chandeliers. Studio Booth reads as a real session behind foamed walls, and Street Cypher brings neon, wet asphalt and handheld energy.

Pick the stage that matches the occasion, not the most expensive-looking one. A birthday in a marble lobby reads as a celebration; a rap battle in a recording booth loses the street tension. If you are torn, render a Quick version on one stage before spending your second free render.

Step 3: Write a short cue instead of a script

The cue is two fields: who is in the video, and why it exists. 'Dani and Alex, two years, terrible at bowling' gives the model a relationship, a milestone and a joke in nine words. That is the entire craft — specificity, compressed.

Avoid vague cues such as 'make it fire' — they tell the model nothing it could not have guessed. Avoid real artists, songs and shows too: the render should be your duo, not an impression.

• Birthday: 'Tasha's 30th, no gifts just bars'.
• Best friends: 'Jo and Chris, ten years of friendship'.
• Promo: 'Rosa's Tacos, two-for-one Tuesdays'.
• Couple: 'Sam and Ali, first apartment together'.

Step 4: Choose a quality tier, and test with Quick first

Quick gives you 8 seconds at 480p and is the fastest way to check whether both faces read correctly. Standard is 10 seconds at 480p — the sweet spot for TikTok and Reels. HD is 10 seconds at 720p and counts as two free renders during the beta, so it is the expensive choice.

Because each IP gets two free renders, order matters: one Quick test to validate the photos, then one Standard or HD render for the version you post. A failed render does not consume a free render, so a technical error is not a reason to panic.

Step 5: Download, caption and post

Download the MP4 as soon as the render finishes. It is vertical 9:16 and watermark-free, so it fits every feed without cropping. Move it to your phone first — native uploads play nicer than a file link.

Caption in one or two lines. Lead with the relationship, not the technology: 'ten years of friendship and this is what AI did with it'. Say plainly that the video is AI-generated, then let the joke carry the rest. If the clip is for a friend, tag them; if it is for a trend, use the trend's sound-free version of the hook and keep the text readable.

• Say it is AI in the caption or on-screen text.
• Tag the other person so the post reaches both audiences.
• Keep captions short enough to read without pausing.

The mistakes that waste a free render

Almost every disappointing render traces back to one of five choices. Fix the photo first, then the stage, then the cue — change only one thing at a time so you know what worked.

• Using a group photo or a screenshot as the face reference.
• Writing a paragraph when a line would do.
• Changing stage and quality at the same time.
• Uploading someone else's photo without asking.
• Posting without disclosing that the clip is AI-generated.