prep

AI Rap Video Photo Prep Checklist

Most AI rap video failures are photo failures wearing a prompt's clothes. A blurry identity, a face that morphs at second six, a duo where one performer always looks "further away" — all of them trace back to the input images. This checklist is the five minutes of prep that saves five re-renders.

Why prep beats prompting

Image-to-video models treat your portrait as ground truth for identity. Anything ambiguous in the photo — soft focus, a tilted head, a busy background, sunglasses — gets resolved by the model's imagination, and imagination drifts frame by frame. A sharp, front-lit, plain-background photo pins the identity; the prompt's job is then only staging and choreography. Run this checklist before touching any template.

The checklist

1. Resolution: at least 512px on the face. The face should fill a meaningful part of the frame; a face occupying 5% of a group photo does not carry an 8-second render.
2. Front-facing, eyes visible. Three-quarter angles work; full profile and heavily tilted heads do not. Eyes open, no heavy shadow across them.
3. Even front light. Daylight from a window or a bright room beats a dim party flash. Harsh side shadows confuse the identity mapping.
4. Plain or blurred background. The model can read background objects as props to render. A clean wall keeps the stage under prompt control.
5. Upper body in frame. The booth performance shows chests and shoulders; a face-only crop forces the model to invent a body, and it will change clothes mid-clip.
6. Glasses: keep them only if they're identity. Clear lenses with no glare usually survive; strong glare or sunglasses usually do not.
7. Similar scale across photos. For a duo, both faces should be cropped to roughly the same size, angle, and distance so neither performer reads as "the far one".
8. One face per photo. Never split a group shot into two uploads — the model keeps the coupling and merges faces.
9. Recent and representative. A photo from ten years ago renders a person the subject won't recognize. Pick something they'd approve of.
10. Permission. Every face in a photo you upload must have agreed to be in the video. No exceptions, including your own group chats.

Copy-ready prep verification prompt

Use this in any image-capable chat model to audit your photos before rendering. It is the one prompt in this wiki that runs against images instead of video:

Audit these portrait photos for an AI rap video render. For each photo, answer strictly:
1. Is exactly one clear face visible? (yes/no)
2. Is the face front-facing with both eyes visible? (yes/no)
3. Is lighting even across the face with no harsh shadows? (yes/no)
4. Approximate face resolution in pixels (width of the face region).
5. Is the background plain or clearly separated from the subject? (yes/no)
6. Is the upper body (chest and shoulders) visible? (yes/no)
7. Any glare on glasses, heavy motion blur, or heavy compression artifacts? (list them)
Then give a verdict per photo: RENDER-READY or FIX FIRST, and for any FIX FIRST photo, name the single highest-impact fix.
Do not edit the images. Do not generate video. Audit only.

How the audit maps to fixes

• Low face resolution → re-crop tighter around the face rather than upscaling; interpolation adds the softness you're trying to remove.
• Harsh shadows → reshoot near a window, or pick a different photo from the same set.
• Two faces in frame → crop to one, or use the second face as the other performer's reference if they consent.
• Glare on glasses → tilt the head slightly in a reshoot, or accept the risk and plan for a re-render.

Common failure modes

• "My render looks nothing like them" → almost always resolution or angle. Re-crop to 512px+ face width, front-facing.
• "The two performers look like different distances away" → scale mismatch between photos; re-crop both to the same face size.
• "The wardrobe keeps changing" → face-only crop; widen to chest-up so the outfit is real reference data, not invention.

FAQ

Does a higher-resolution photo always help?

Up to a point. What matters is the face region's sharpness, not the file's megapixels. A 12MP photo where the subject is small in frame loses to a 2MP selfie where the face is large and sharp.

Can I prep photos with AI enhancement tools?

Light cleanup (denoise, exposure) is fine. Heavy face "beautification" or upscalers rewrite identity features — the render will then faithfully animate a face that isn't quite theirs.