APT AI Video Prompt: The MiniMax H3 and Seedance Settings That Work
The exact prompt, reference clip and settings for making the Rosé & Bruno Mars APT. duo video yourself with MiniMax H3 or Seedance 2.0 reference-to-video, and why a short prompt beats a long one.
The two-photo APT AI video is a reference-to-video job: one short clip from the official "APT." video supplies the motion, camera and timing, and two photos supply the people. Our one-click APT generator sets all of that up for you. This guide is for anyone who wants to run it by hand in MiniMax H3, Seedance or another reference-video model, and it uses the same prompt and settings our generator sends.
What you need
- A model that takes two character images and a reference video in one generation. MiniMax H3, Seedance 2.0 (Mini, Fast and Pro), Seedance 2.5 and Wan 3.0 all do. Motion-control tools that accept only one character image won't work for a duo: one of the two performers is left unchanged.
- A short reference clip from the official video, 5 to 15 seconds long, with both performers in the frame.
- Two photos, one person per photo, front-facing, upper body visible.
Pick the reference clip carefully
The clip matters more than the prompt. The model copies everything about it: framing, cuts, who stands where. What we looked for when we cut ours:
- Both performers in frame for most of it. Our clip starts at 0:35.56 of the official video, on a shot of the two of them together, and runs through the drum-kit section and the first dance and floor shots.
- Left and right stay put. In most of the video Bruno Mars is on the left and ROSÉ on the right. At 0:51.7 there's a sunglasses close-up where they swap sides, so our 15-second clip ends at 0:50.56, just before it. There's still one 0.8-second close-up (about 11 seconds into the clip) where the positions flip, which is why a 15-second result can briefly put a face on the wrong performer.
- Same length as the output. Cut the reference to exactly the duration you'll request (5, 10 or 15 seconds), so the model isn't stretching or truncating it.
- Landscape. The video is 16:9. A vertical crop cuts one performer out.
We scale the excerpt to 1280 × 720 at 30 fps and keep its audio track, because the sound helps the model time mouth movements to the song.
The prompt
This is the full prompt our generator sends. It's deliberately short:
Replace both performers in @Video1.
Replace the performer on the viewer's left with the person in @Image1.
Replace the performer on the viewer's right with the person in @Image2.
Keep both reference identities consistent throughout the video, including their faces and hairstyles. Neither original performer should remain.
Preserve the original background, framing, movements, timing, and audio.
Upload the photos in that order: @Image1 becomes the left performer (Bruno Mars's spot) and @Image2 the right (ROSÉ's spot). The reference-token syntax differs between tools; use whatever your tool calls the first video and the first and second images.
Why not describe the scene?
It's tempting to describe the pink studio, the leather jackets, the drum kit and the choreography. Don't. When we tested this on our two-person Hotel Lobby template, a long descriptive prompt with a separate audio reference made the motion drift and gave both people the wrong gestures. We measured how closely the result's motion followed the reference clip: the short prompt with the clip's own audio scored a motion correlation of 0.85, the long prompt 0.74. The reference clip already contains the scene; the prompt only needs to say who replaces whom.
Settings
| Setting | Value |
|---|---|
| Mode | Reference to video (two images + one video) |
| Duration | The length of your reference clip: 5, 10 or 15 s |
| Aspect ratio | 16:9 |
| Resolution | 720p (MiniMax calls it 768P) |
| Generated audio | Off: lay the reference clip's audio back in afterwards |
| Watermark | Off, if your tool offers the option |
Put the original audio back
Generating with audio off and adding the reference clip's track afterwards keeps the song exactly as it was. With FFmpeg, where generated.mp4 is the model's output and reference.mp4 is your reference clip:
ffmpeg -i generated.mp4 -i reference.mp4 -map 0:v -map 1:a -c:v copy -c:a aac -shortest apt-ai-video.mp4
Our generator does the same thing, plus scaling to 1280 × 720.
Which model to use
- Seedance 2.0 Mini: the cheapest way to test a pair of photos. Run 5 seconds first.
- MiniMax H3: a good balance of price and likeness for two faces.
- Seedance 2.0 Pro and Seedance 2.5: more detail, at a higher cost.
- Wan 3.0: a different look; worth trying if a pair doesn't hold up on the others.
On TryH3 you can run the manual workflow in reference to video, or skip it entirely: the APT AI Video Generator already has the reference clip, prompt, audio and settings loaded, so you only upload two photos. Per-video prices for every model are on that page.
Common problems
- One of the original performers is still there: say "Replace both performers" explicitly, and check that both photos were attached as character references, not only one.
- The two faces blend: use one person per photo and front-facing, evenly lit shots. Group photos and sunglasses cause most blends.
- Faces swap sides partway through: the reference clip has a shot where the performers trade places. Trim it out, or keep the clip at 10 seconds or less.
- Moves drift away from the original: shorten the prompt. Scene descriptions compete with the reference clip.
For a non-technical walkthrough, see how to make the APT AI video from two photos. To compare the one-click tools, see the APT AI video generator comparison.
The reference clip is a short excerpt of the official "APT." video. TryH3 is an independent tool and is not affiliated with or endorsed by ROSÉ, Bruno Mars, THEBLACKLABEL or Atlantic Records.