MiniMax H3 Character Consistency: Keep the Same Face Across Every Shot

Why the face drifts between shots

This guide is for anyone making a series of shots with the same person in them: short dramas, story videos, product videos with a recurring host. It applies to MiniMax H3, the open-weight video model from MiniMax that is also available in the Hailuo app, and the approach carries over from earlier Hailuo subject-reference models.

With only a first frame, the model knows one view of the person: whatever angle that image shows. As soon as the shot turns to a profile, the person walks away, or the camera pushes into a close-up, the model has to invent the missing views, and every shot invents them differently. That is why the same person looks like a cousin of themselves three shots later.

Reference photos fix this by giving the model the views it would otherwise guess. On XXT, every job with a person on screen should carry reference photos or a saved character, first-frame and lip-sync jobs included. Text-to-video has no image, so don't use it for a recurring person; use a free shot instead.

Step 1: Prepare the photos

• 3–5 sharp close-ups of the same person. Fewer than 3 can be unstable; one job takes up to 8 images (7 when you also give a last frame).
• At least one straight-on front photo. It is the main identity cue.
• Add three-quarter left and right and a profile if you have them; full-body front and back photos help for full-body shots.
• A plain background, even light, nothing covering the face.
• Don't use a small face cropped out of your first frame, or an AI-composited medium shot: they blur when the model enlarges them, and the blur becomes part of the identity.

Step 2a: One-off, attach photos to a single job

Label each photo with its angle: front, three_quarter_left, three_quarter_right, side, full_front or full_back.

h3 submit --mode ref2v --aspect 16:9 --seconds 8 \
  --ref front=mia_front.jpg --ref three_quarter_left=mia_left.jpg --ref side=mia_side.jpg \
  --prompt "A medium shot frames the woman sitting at a café window on a rainy afternoon, both hands around a cup of coffee. She looks out at the street, then turns toward the camera and smiles. The camera pushes in with small amplitude at slow speed in one continuous shot." \
  --wait --out ./out

The same --ref flags work in image-to-video (i2v) and lip-sync, so a shot that starts from your own first frame still keeps the face.

Step 2b: Recurring person, save a character

When the person appears in more than one shot, create a character once and reuse it. A character holds 1–8 photos and an optional voice; photos and voice stay in your account until you delete them.

h3 character create --name Mia \
  --photo front=mia_front.jpg --photo three_quarter_left=mia_left.jpg --photo side=mia_side.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

Without a recording, the platform designs a voice from --voice-description. Listen to it (h3 character voice Mia gives a link) and redesign until it fits. Then pick the character in any shot:

h3 submit --mode ref2v --aspect 16:9 --seconds 8 --character Mia \
  --prompt "Mia walks into a quiet bookstore in the evening, runs her fingers along a shelf and pulls out a worn paperback. A medium close-up, the camera tracks her slowly from the side." \
  --line "Mia:opens the book and smiles=I've been looking for this one for years." \
  --wait --out ./out

Characters work in free shots (ref2v), image-to-video (i2v), lip-sync (one character) and promo videos. Write the character's name straight into the prompt; the platform links it to the photos. The full commands are in the characters reference.

Step 3: Pick the right angles for each shot

By default the platform attaches all of a character's photos. For a specific shot, fewer matching angles beat many: --character Mia=front,side uses only her front and profile photos. Always include front.

Step 4: Write the prompt so it doesn't fight the photos

• Don't describe the face, hair or build. The photos set them; words about looks compete with the photos. Spend the words on action and camera.
• Small faces in wide shots blur. Include a sharp close-up taken at the angle the person faces in that shot (front when facing the camera, side in profile), and distant faces stay clear. Without a matching photo, show the face only in close-ups and medium close-ups, and the person from behind in wide shots.
• Say there is one person when there is: the same single woman throughout. Otherwise turning, walking past the camera or leaving and re-entering the frame tends to produce a second copy of her.
• In close-ups, don't mention clothes outside the frame (trousers, shoes), or the model widens to a full-body shot to show them. We tested this.

The general prompt rules are in the MiniMax H3 prompt guide.

Clothes and location

• Clothes follow the prompt; the photos only set face, hair and build. To keep the clothes from the photos in a free shot, add --wardrobe Mia and don't describe her clothes.
• Lock a location across shots by attaching the same scene image every time, for example --ref "other:living room=living-room.jpg", and just write "living room" in the prompt. Scene and prop images share the 8-image budget with the character photos.

Voices stay consistent too

A character's voice is used every time they speak (--line "Mia:…=…"), at most 3 speakers per shot. In a film with several people, make their voices differ in at least two of age, timbre and pace, so viewers can tell who is speaking. Auditions for a new voice are free, 20 per day. For a common role you can also pick a ready-made voice template (child, teen, young, middle-aged, elderly, strong persona, host; one woman and one man each; list them with h3 character voice-templates, use one with --voice-template), free and not counted as an audition.

When it still goes wrong

What it costs

Reference photos and characters add nothing to the price. Video is billed per second: $0.009 per second at night (22:00–10:00 Singapore Time) and $0.0135 during the day, so an 8-second 768P shot submitted at night is 48 credits ($0.072). 1080P costs double.

Videos, songs and uploaded files are kept for 8 hours after the job ends, then deleted automatically. Please download them in time. Character photos and voices are the exception: they stay until you delete them.

FAQ