troubleshoot
Fix Face Mix-ups in AI Rap Video
The render came back and something's wrong with the faces. Don't rewrite the whole prompt — diagnose. Face failures in AI rap video renders have about seven shapes, each with a specific cause and a one-change fix. This page is the map: find your symptom, apply its fix, re-render once.
Symptom → cause → fix
| Symptom | Likely cause | Fix (one change) |
|---|---|---|
| Faces swap sides mid-clip | Left/right assignment was stated once, early, then lost as motion ramps up — or the two references look too similar. | Restate "fixed sides for the whole clip" late in the prompt, not just the top. If both outfits are close in color, re-crop references with stronger contrast (see the self-duo contrast dials). |
| One face goes soft or "melts" after a few seconds | That performer's reference is lower resolution or smaller in frame, so the model is interpolating identity it can't see. | Re-crop that photo so the face is ≥512px and brighter. Don't upscale — interpolation adds the same softness back. |
| A third face appears briefly | The model invented a background onlooker when the stage description was vague during a camera move. | Lock the camera (drop to a static shot) and add "no other people" plus the stage's "dark backdrop, empty stage" language. Then re-add movement one piece at a time. |
| Faces fuse during a lean or embrace | Contact choreography — the couple lean, a high-five hold — puts the two identity regions adjacent, and the model blends them. | Keep contact shoulder-level and brief ("one brief shoulder-to-shoulder lean"), or remove contact entirely. Distance is the anti-fusion tool. |
| Both performers rap every line | The call-and-response roles weren't labeled per line, so the model plays the audio on both faces. | Rewrite the audio direction as a LEFT:/RIGHT: script (one line each, alternating) instead of a lyric paragraph. |
| A face changes but stays plausible | Wrong-person substitution: the slot order in the tool didn't match the prompt's "photo 1 = left" assignment. | Check the upload order first — it overrides wording in most tools. Swap slots to match, don't rewrite the prompt. |
| The whole pair morphs (self-duo) | Identical faces on both sides with insufficient outfit contrast — the model can't hold "two" of the same person. | Push outfit colors to maximum separation, add hairstyle/headwear difference, and keep a visible gap between the performers. Full dial list in the self-duo template. |
The layered-fix ladder
When a symptom has multiple candidate fixes, change layers in this order — cheapest and most causal first:
- Layer 1 — Slot order. Check which reference went into which upload slot. This silently overrides wording in most tools and explains "wrong person" renders instantly.
- Layer 2 — Photo quality. Resolution, lighting, crop. A better input fixes more than any wording change (see the prep checklist).
- Layer 3 — One prompt clause. Add or reinforce the single sentence that names the failure ("fixed sides for the whole clip", "the microphone is never held"). Never rewrite the whole prompt.
- Layer 4 — Negative terms. Add the matching blocklist term if it's missing (start from the negative library).
- Layer 5 — Choreography. Cut the riskiest gesture (high-five, lean, spin). Simplify before you escalate.
- Layer 6 — Duration. Drop from 12s to 8s. Most morphing is a late-clip decay; a shorter clip sidesteps it.
If a failure survives all six layers, the render request is at the edge of what the model can hold — usually three-plus faces or an ambitious continuous camera move. Split the scene into two renders and cut them together.
The 60-second review checklist
Before posting — or before deciding anything is wrong — watch the clip once, frame by frame through the midpoint:
- Left performer stays left, right stays right, from first frame to last.
- Both faces sharp at the midpoint (the most common decay point).
- Exactly the number of people you cast — no background cameos.
- Mouths move only on the rapping performer's lines.
- Wardrobe and sides match between the first and last frame (loop integrity).
Common failure modes (meta)
- Fix-shuffling: changing three things per re-render, so you never learn what worked. One change, one render, one observation.
- Prompt-blaming: most identity failures are input failures; the checklist above starts at slots and photos for a reason.
- Escalating choreography: adding moves to "make it better" while diagnosing. Freeze the choreography until the faces are stable, then decorate.
FAQ
Can I fix a face swap in post instead of re-rendering?
Small fixes (one soft frame, a flicker) can be trimmed or masked. An identity swap can't — the wrong face is baked through the clip. Diagnose and re-render; it's faster than it sounds once you're fixing one layer at a time.
Is there a prompt that just never fails?
No — but the failure rate drops an order of magnitude with sharp photos, fixed-side wording, one camera move, and the negative block. That's what every template on this wiki ships with.
Related AI Rap Video templates
AI Rap Video Photo Prep Checklist
Choose and crop photos so faces stay recognizable in an AI rap video.
AI Rap Video Duo Prompt (Two Photos)
Build a two-person AI rap video from two separate portraits.
Negative Prompts for Cleaner AI Rap Video
Reduce face swaps, blur, NSFW, and text artifacts in AI rap video renders.
Self-Duo AI Rap Video (Two Outfits)
Use two outfits of the same person as left and right performers.