Workflow guide
How to make an AI rap video
A prompt-first workflow that works in any image-to-video model: prepare photos, lock left/right roles, write original bars, render an 8-12 second clip, then review it for the identity drift that ruins most first attempts. No credits, no upload queue — this wiki supplies templates, your model supplies the render.
Before you start: three decisions
- Duo or self-duo? Two people from two photos, or one person from two outfits. Both work; the self-duo template handles the contrast rules for one person.
- Vertical or landscape? TikTok, Reels, and Shorts want 9:16. YouTube handles 16:9. The vertical template bakes the framing into the prompt.
- What is the clip about? A birthday, a reunion, a couple milestone, a pet. The theme drives the original bars, so decide it now.
The 10-step workflow
- Pick your performers and get permission. One face per photo, and a real yes from every person in it. Duo clips from two separate portraits are the cleanest start; two outfits of one person work too.
- Prep the photos. Front-lit, eyes visible, upper body in frame, plain background. Crop each face to roughly the same size before uploading.
- Choose the frame. Vertical 9:16 for TikTok, Reels, and Shorts; 16:9 for YouTube. Decide before you prompt — retrofitting framing after a render wastes a generation.
- Write original bars. A 4-line hook plus an 8-bar verse about your subject. Never paste commercial song lyrics — they get clips muted or taken down.
- Assemble the main prompt. Combine the lyric direction with stage and camera language: booth stage, hanging mic, locked two-shot, slow push-in, loop ending.
- Add negative prompts. Paste the negative block for your template: identity swap, merged faces, on-screen text, logos, NSFW.
- Render in your own model. Drop the two photos into your image-to-video tool of choice, set 8-12 seconds at 9:16 or 16:9, and run the render. This wiki does not render — any Wan-style or Seedance-style workflow works.
- Review for identity drift. Watch the full clip once. Check that the left performer stays left, both faces stay sharp, and no third face appears mid-clip.
- Fix and re-render if needed. Most failures trace back to photo prep or ambiguous left/right wording. Apply the matching fix, not a full prompt rewrite.
- Loop and post. End the clip on a clean two-shot hold so it loops seamlessly, then caption it. Keep the audio original and the faces permissioned.
What each step feeds
The steps chain into each other. Photo prep feeds identity stability, left/right wording feeds side consistency, original bars feed the audio direction, stage language feeds the visual, and negatives protect all four from common render failures. If a render disappoints, diagnose by layer instead of rewriting the whole prompt — the face troubleshooting guide maps symptoms to the layer that caused them.
Which model should you use?
This wiki is deliberately tool-agnostic. The templates are written in plain descriptive English, so they port between Wan-style and Seedance-style image-to-video workflows with small wording shifts. The Wan vs Seedance prompt guide covers exactly what to change — reference image handling, duration defaults, and where the negative block goes in each style.
FAQ
How long does the whole workflow take?
First run: about 20 minutes of prep plus your model's render time. Re-runs are faster because the photos and prompt structure are already done — usually one wording tweak and one re-render.
Can I make an AI rap video with just one photo?
Yes, as a solo performer, or as a self-duo if you have a second photo in a different outfit. The prompt changes one block: assign the single reference and drop the left/right trade.
Why not just use a trend generator site?
Trend generators lock you to their stage, their credits, and often a specific commercial audio. Our compliance comparison breaks down why prompt templates keep you in control and on the safe side of music and likeness rights.