How to Make the Burning Bridges AI Video, Step by Step
This is the character swap method: the one that keeps the real room, crowd and camera movement and replaces only the lead. If you just want a quick clip from one photo, use the quick start on the home page.
- Trim 8–20 seconds of the party scene around the wave.
- Use Object Swap, not Restyle. Restyle repaints the whole room.
- Tell the tool exactly who replaces whom and to keep everything else unchanged.
- Plan for 3–5 takes.
What you need
- A short clip of the scene that you are allowed to use. The official video belongs to OVO and Republic Records; this site does not host or supply it.
- A clear photo of each new person or character. One photo is enough if you only replace the lead. For characters, guides suggest front, three-quarter and full-body views.
- A character swap tool. The documented versions used Higgsfield Genjutsu, which accepts a reference video of roughly 4–30 seconds and up to 30 reference images. Seedance 2.5 reference-to-video and Kling Motion Control are named as alternatives.
The swap, step by step
- Trim the source
Cut 8–20 seconds from the party section, centred on the wave at about 2:58 and the cup raise. Keep the handheld movement; it is what makes the result read as the original shot. - Prepare reference images
One clear, evenly lit photo per person. Avoid harsh flash and heavy filters. If only part of the outfit is visible, the model invents the rest, so use a photo that shows what you want to wear. - Choose Object Swap
In Genjutsu, Object Swap changes the selected person and keeps the room. Restyle changes the look of the whole frame, which is the wrong mode for this trend. - Map who replaces whom
Point the lead to your photo. Optionally map the friend at frame-right to a second photo. Leave everyone else as they are. - Write a short prompt
Say which person in the clip becomes which image, ask it to match the pose, wave, cup and position, and tell it to keep the room, lighting, other people, camera movement and cuts unchanged. Ready-made versions are on the prompts page. - Generate and review
Check the face, the cup grip, the hands, the background guests and the camera continuity. Rerun until one take holds up. - Export and post
Crop to 9:16. If the file has no audio, add the official sound in the app and slide the clip so the hand goes up as the line is sung. Turn on the AI-generated label.
Common failures and fixes
| Problem | What to try |
|---|---|
| The cup merges with the hand | Name the cup hand in the prompt and rerun |
| The cup switches hands | State which hand holds it in both the still and the video prompt |
| Background faces drift or change | Add a line that locks the other people and the room as unchanged |
| Faces trade places with two or more swaps | Swap fewer people, or map each one by position in the frame |
| Skin picks up too much of the room's colour | Use a neutral, evenly lit photo without filters |
| The whole room looks repainted | You are in Restyle mode. Switch to Object Swap |
| Lip sync breaks on the key line | Keep the hand over the eyes, not over the mouth |
| A motion tool follows the wrong person | Kling Motion Control handles one character and follows the largest figure in frame. Use a swap tool for multi-person shots |
Collected from the Top5Apps, Starrd, EAST.IM and Media.io guides.
The advanced route: depth map and beat list
Starrd's guide describes a method that avoids feeding the raw clip to the model, because the original performer's face and outfit can carry over:
- Make a depth map. Strip the audio and run the take through a depth-estimation model such as Depth Anything V2. That keeps the movement and camera motion without the original person's look.
- Generate a still. A vertical 9:16 film still of yourself at the party, with the cup hand named explicitly.
- Write a beat list. Watch the original at two frames per second and log each move with its timestamp, for example “16–17.5s: right hand over the eyes, cup up beside the face”.
- Generate. Give the still, your photo, the depth video and the beat list to a motion-transfer model (the guide recommends Seedance 2.5), then add the song back at the end.
A depth map alone gives rough motion. The signature moves only land when they are written into the prompt with timings.
Before you post
- Use your own face or people who agreed. Do not use photos of strangers or children.
- Turn on the AI-generated label. TikTok, Instagram and YouTube require it for realistic AI content.
- Do not present the result as a real scene from the official video.
- Parody posts are common, but that does not make the clip yours to sell or monetise.
FAQ
How many tries does it take?
Guides report three to five runs before one looks right.
How long should the clip be?
Trim 8–20 seconds. Genjutsu accepts reference videos of roughly 4–30 seconds.
Can I swap more than one person?
Yes, but the more people you replace the more often faces trade places. Start with the lead only, then add the friend at frame-right.
Do I need the original clip?
Only for the swap method. One-photo generators and prompt-only image-to-video do not need it, though the result will resemble the scene instead of matching it.
Can I do it on my phone?
The one-photo templates run in a mobile browser. Trimming a clip for the swap method is easier in a video editor, on a phone or a computer.
Unofficial, independent guide. Not affiliated with, endorsed by or connected to Drake, OVO Sound, Republic Records or any AI tool mentioned here. We do not host the music video, the song or any copyrighted footage.