MiniMax H3 Lip Sync: Make a Photo Sing or Talk to a Voice Track

What you need

• A first frame: the person who will sing or talk, ideally a frontal close or medium-close shot, short side 720 px or more. It sets the scene, clothes and light.
• A voice track: clean vocals, ideally without music, 2–60 seconds. MP3, WAV, M4A, AAC, FLAC or OGG. If your track has music, separate the vocal first; if it is longer than 60 seconds, cut it and make several clips.
• Reference photos or a saved character for the same person, so the face holds when they turn or the camera moves. See character consistency.

The command

h3 submit --mode lipsync --first-frame girl_river.png --audio verse1.wav --seconds 8.4 \
  --ref front=face_front.jpg --ref three_quarter_left=face_left.jpg \
  --prompt "A medium close-up frames the young woman standing by the river at dusk, facing the camera. She sings with gentle, expressive facial movement, her right hand resting lightly on her chest, her hair moving in the evening breeze. The camera pushes in with small amplitude at slow speed in one continuous shot." \
  --dialogue "The evening wind is on the river, and I'm still waiting for you." \
  --sound "Soft wind over the river, water lapping against the bank, distant birds." \
  --wait --out ./out

--seconds is the length of the voice track. With a saved character, use --character Mia instead of --ref: the platform attaches all of her photos, and the mouth still follows your voice track.

Write the dialogue word for word

• --dialogue must match the voice track word for word, punctuation included, or the lips drift out of sync.
• Keep it in the language that is sung or spoken: Chinese lyrics stay Chinese. It is never translated.
• End it with a full stop.

Describe the performance, not the lyrics

The description covers how the person sings or talks: expression, gestures, small body movements and the camera. Don't repeat the lyrics there. For instrumental gaps, add keeps relaxed closed lips during instrumental sections and sings only when vocals are audible.

The platform already adds the instruction that every syllable is clearly articulated, so don't write it yourself.

Words that freeze the mouth

Avoid restrained, subtle, minimal, almost still and similar words. In our test with the same seed, image and vocal, these words left the mouth barely moving, just showing teeth; without them the mouth opened and closed clearly. Describe the expression and gestures instead and let the voice drive the mouth.

Framing that syncs best

• Frontal faces in close shots are the most stable.
• Profiles, a hand or hair covering the mouth, wide shots, several people in frame and fast head turns make sync worse.
• If the face is small in your image, crop the first frame closer before submitting.
• Crop the first frame to your aspect ratio (9:16, 16:9 or 1:1); otherwise the platform center-crops it.

What to expect

MiniMax H3 generates sound and picture together; it is not a forced lip-sync tool, so the lips occasionally mouth something slightly different. For important shots, generate two takes (--segments 2, double credits) and pick the better one.

A music video is several lip-sync clips plus a few shots without singing. Each clip is billed per second like any video: $0.009 per second at night and $0.0135 during the day. You can also make the song itself on XXT: 2 tracks per run, 100 credits per track.

Videos, songs and uploaded files are kept for 8 hours after the job ends, then deleted automatically. Please download them in time.

FAQ