Two-Person Dialogue in MiniMax H3: Fix the Wrong Character Speaking and Burned-In Subtitles

Why the wrong person speaks

MiniMax H3 generates the picture and the sound together. With two faces in frame and a line that isn't tied to anyone, the model has to guess which mouth should move, and it guesses wrong often enough to ruin a scene: the listener mouths the words, both people talk at once, or the line comes out in the other person's voice. The same applies to the Hailuo app and to most AI video models that generate dialogue.

Three things fix most of it: give each speaker a distinct identity and voice, tell the model exactly who says which line, and frame the shot so the speaker is the obvious one to talk.

Step 1: Make each speaker a character with a voice

A character holds a few face photos and a voice. Every speaker needs a voice; a character without one can be on screen but can't talk. Make the two voices differ in at least two of age, timbre and pace, so viewers can tell them apart even when the picture is ambiguous.

h3 character create --name Mia \
  --photo front=mia_front.jpg --photo three_quarter_left=mia_left.jpg \
  --description "a woman in her late twenties with shoulder-length black hair" \
  --voice-description "a woman around 28, warm and slightly breathy, mid-low pitch, medium pace, relaxed tone" \
  --language English
h3 character create --name Wang \
  --photo front=wang_front.jpg --photo side=wang_side.jpg \
  --description "a man in his forties with short grey hair and a beard" \
  --voice-description "a man in his forties, deep and slightly husky, low pitch, slow pace, gentle tone" \
  --language English

Listen to each voice before you shoot (h3 character voice Mia gives a link). More on photos and voices in character consistency.

Step 2: Write one line per speaker, in order

Pass each line with --line, in speaking order: --line "Name=text", or --line "Name:action=text" to say what the character does while speaking. This is the single most effective fix: the platform ties each line to that character's face and voice.

h3 submit --mode ref2v --aspect 16:9 --seconds 8 \
  --character Mia --character Wang \
  --prompt "A medium shot frames a small kitchen at night, warm light from a pendant lamp over a wooden table. Wang, in a dark blue sweater, stands at the stove stirring a pot; Mia, in a grey hoodie, sits at the table with a laptop. The camera holds a static shot." \
  --line "Mia:looks up from the laptop=Is dinner ready yet?" \
  --line "Wang:turns from the stove and smiles=Five more minutes." \
  --wait --out ./out

• At most 3 speakers and 10 lines per shot.
• Write the action without a subject: turns from the stove and smiles, not he turns….
• End every line with a full stop or other closing punctuation, or the speaker tends to keep mumbling.
• You can also put lines in quotes in the prompt with the speaker's name before them, and the platform splits them out. Splitting them yourself with --line is more accurate.

Step 3: Put the camera on whoever is speaking

• Frame the speaker. A speaker facing the camera with the listener in profile or from behind leaves little room for doubt. Two frontal faces side by side is the hardest case.
• Tie lines to actions, not time points. "Mia looks up and asks…" works better than "at 3 seconds Mia says…".
• Give the listener something silent to do: listening, nodding, stirring a pot. Outside the lines, avoid words like talk, chat or argue, or extra voices appear.
• For an important exchange, shoot it as separate shots: one clip per line, each framed on its speaker, then join them. This is how film dialogue is cut anyway, and it removes the guesswork. See longer videos for joining clips. Even a one-line clip sometimes cuts to the listener on its own, so check every clip.
• Watch for unrequested close-ups. In our tests, the usual way a line lands on the wrong person was not a mouth swap: the model cut on its own to a close-up of someone who wasn't speaking, sometimes that character's reference photo pasted in whole, and the next line played over it. Two people in a static two-shot rarely had this problem; three people at one table had it in more than half of the takes.
• Three people in one shot: give each a position. Add --position for every character (--position Mia=the young woman on the left); the prompt then introduces everyone by position, names each speaker by position, puts the lines in order and has the speaker's lips close after each line while the others listen. At our three-person dinner table, the same seeds put 36 of 40 lines on the right mouth with positions and 32 of 40 without. In the web studio, fill in the position box (站在哪) under each picked character. Details in the characters reference, section 6.

Why lines show up as subtitles, and how to stop it

Sometimes the model burns the dialogue into the picture as captions. We tested this on two-person dialogue: with a scene of only a sentence or two, the lines often appeared as subtitles; with the scene written in full and a closing action after the last line, they didn't.

• Describe the scene in full: shot size and camera position, the location and set dressing, the lighting, where each person is, what they wear and what they are doing before they speak.
• After the last line, add a silent closing action, such as Mia closes the laptop and stands up.
• In a free shot (ref2v) with lines, the platform expands a short scene and adds a closing action for you, but what it adds may not be what you want; writing it yourself is safest.
• Don't write "no subtitles" or "no text". There is no negative prompt in MiniMax H3, and naming the thing tends to pull it in.
• If subtitles still appear, resubmit with a different --seed, or crop or cover them in editing.

Fit the lines to the duration

Lines that are too long for the clip make mouths race and words blur. For Chinese, keep to about 7 characters per second; for English, a short sentence per speaker in an 8-second shot is plenty. If the conversation is longer, split it across shots rather than packing one clip.

Dialogue from your own first frame

If you already have an image of the two people, use image-to-video with both characters and the same --line flags. The first frame sets the scene, so the platform doesn't expand it; describe the action and camera yourself.

h3 submit --mode i2v --first-frame kitchen.png --aspect 16:9 --seconds 8 \
  --character Mia=front,three_quarter_left --character Wang=front,side \
  --prompt "Mia sits at the kitchen table with her laptop while Wang stands at the stove. The camera pushes in with small amplitude at slow speed toward Mia. After the last line, Wang turns back to the stove and keeps stirring." \
  --line "Mia:looks up from the laptop=Is dinner ready yet?" \
  --line "Wang:glances over his shoulder=Five more minutes." \
  --wait --out ./out

For full control over who says what and when, including voice-over, you can write the complete MiniMax H3 prompt yourself with --raw; the format is in the prompting reference.

Quick fixes

What it costs

Characters, voices and lines add nothing to the price; voice auditions are free, 20 per day. Video is billed per second: $0.009 at night (22:00–10:00 Singapore Time) and $0.0135 during the day. For an important dialogue shot, generate two takes (--segments 2, double credits) and pick the one where the right mouth moves.

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

FAQ