Putting Your Face Into a Music Video Scene — Three Methods That Solve Different Problems

There's a trend of taking a scene from a famous music video and swapping in your own face. I wanted to know how people make these. There turned out to be three methods, and I ran all three. One got stuck at the payment step.

The short answer: most of those videos swap only the face. Replacing the whole person is a different method, and editing the first frame to generate a new clip gives you a similar new scene, not a remake. After reading this you can pick the method that matches the result you want.

The ground rules first. The face you put in must be yours or one you have consent for. Publishing something that looks like a real person is impersonation and deepfake territory. Label it as AI-altered when you post. The original MV and its music are copyrighted, so on YouTube you may hit Content ID.

The three methods at a glance

Setup: M2 Max 64GB, an 8-second clip at 1280×720.

1. FaceFusion: original scene kept, original hair too

hyperswap_1a_256 with pixel boost 512, gfpgan_1.4 enhancement, box occlusion mask. 200 frames in 4 to 8 minutes, zero cost.

Even the parts where a cigarette, smoke or a tattooed hand cover the face looked natural. One trap. The default --face-selector-mode reference treats frames where the angle changed as a different person and skips them, so the original face popped back at 1.2 to 1.6 s and 2.4 to 2.8 s. --face-selector-mode one --face-detector-score 0.35 fixed it.

And the hairstyle stays the original. Put in someone with long hair and they come out with a shaved head. I wrote up why in why face swap can't change your hair.

3. Nano Banana Pro + Veo: an homage, not a remake

The models already differed at the first-frame face swap. Gemini 2.5 Flash Image ignored the instruction twice and returned the original image, and before I specified 16:9 it even cropped it square. Gemini 3 Pro Image (Nano Banana Pro) got it in one go.

Then Veo rejected that frame.

Sorry, we can't create videos from input images containing celebrity or their likenesses.

The face was changed, but the framing, tattoos and outfit were straight from the MV, so my guess is it tripped the celebrity filter. Generating a fresh "scene with a similar mood" without the MV frame went through, and I got 8 seconds of video with sound.

Even with "locked-off static camera, no zoom" in the prompt, both runs pulled the camera back into a wide shot. This route is an homage, not a remake. Trend videos that look slightly off from the original were probably made this way.

2. Runway Aleph: 403 with zero credits

The script was ready, but the upload step returned 403: At least one credit purchase is required for this account to use ephemeral uploads. An account that has never paid can't even upload.

The local alternative, Wan 2.2 Animate, looked like it would take tens of minutes for 8 seconds on a Mac, so I didn't run it. Method 2 has no result yet. Whether Aleph blocks original MV frames the way Veo does is something I'll check after adding credits.

Why people use local tools and apps

Generative APIs (Veo) have celebrity filters. Local tools (FaceFusion) don't. That may be the real reason people who want the original scene end up on FaceFusion or face-swap apps. No filter also means the person using it has to hold the line themselves.

My rule of thumb: keep the original acting and change only the face -> 1. Hair and clothes too -> 2. Just the vibe, different from the original is fine -> 3.

If you're adding AI to video production and not sure which route fits, you can ask negabaro (contact).