How to Create AI Portrait Videos with MiniMax H3 (2026 Guide)

Quick answer

Turning a portrait into video sounds like a small job. The frame already has a person, a setting, and a mood. Portrait animation still punishes small mistakes. A tiny change to the eyes can alter the subject's identity. A large head turn can reshape the jaw. An ambitious camera move can make the background slide while the face stays strangely fixed.

Start small: use a clean, motion-friendly image, describe one short performance, generate a brief clip, and inspect the face before polishing anything else. MiniMax H3 is useful here because it accepts visual references and supports short, high-resolution video generation. It is a general video model, though, not a dedicated talking-avatar service. Treat it as a tool for cinematic portrait motion rather than a promise of perfect speech animation.

This guide shows the complete workflow in the current MiniMax H3 image-to-video generator. The prompt patterns below are practical starting points, not benchmark results.

What makes a portrait video work

A good portrait clip does more than move a still image. It preserves the face while adding a believable change in attention, expression, posture, light, hair, clothing, or background. The viewer should notice the person before noticing the generation.

That usually means keeping the motion modest. A blink, a small breath, a slight shift of the eyes, or fabric reacting to a breeze can make a still portrait feel alive. Five unrelated actions packed into the same shot usually do the opposite. The model has to reconcile too many changes, and identity is often the first thing to drift.

Portrait video also differs from a talking-head avatar. Avatar products are commonly designed around a script, voice, and lip synchronization. MiniMax H3 supports multimodal references, including audio in supported reference workflows, but a portrait image plus a paragraph of dialogue is not the same as a dedicated presenter pipeline. If exact speech timing is the requirement, plan for a separate voice and lip-sync stage. If the goal is a cinematic reaction, fashion moment, creator opener, character beat, or atmospheric profile shot, the workflow here is a better fit.

Pick the right H3 workflow

MiniMax's current video documentation describes three broad paths: text-to-video, first-frame or first-and-last-frame image-to-video, and reference generation. For a portrait, the starting image usually matters more than a long written description. Use the first-frame route when your uploaded portrait already establishes the person, wardrobe, composition, and lighting.

Reference generation is useful when identity or other assets need to guide a new composition. It gives the model more freedom, so the result may not preserve the exact crop or scene from the source. First-and-last-frame generation is better suited to a planned transition where both endpoints matter. It is not automatically safer for facial identity, because the model still has to invent the motion between two fixed moments.

For a first portrait attempt, keep the decision simple:

• Use a first frame when you want to animate the supplied composition.
• Use reference inputs when the person should guide a different shot.
• Use first and last frames only when the ending image is essential to the story.
• Use text-to-video when a specific source portrait is not required.

The public generator exposes image, reference-image, video, and audio input areas along with controls for aspect ratio, duration, generation mode, and resolution. The exact choices available can change, so read the labels in the live interface before starting.

Prepare a portrait that can survive motion

Start with an image that gives the model a clear face and enough visual information to infer a body. Sharpness matters, but simple structure matters more. A crisp portrait with hair covering both eyes is harder to animate than a slightly softer image with a readable gaze, jawline, shoulders, and background.

Choose an adult portrait with these traits:

• The eyes and major facial features are visible.
• The face is not cropped at the chin, forehead, or cheek unless the crop is intentional.
• The shoulders follow a plausible pose.
• Hands are either fully visible or outside the frame. Half-cut fingers invite reconstruction errors.
• Hair edges are readable against the background.
• Lighting has a clear direction rather than several conflicting sources.
• The background includes only elements you want the model to interpret.

A close portrait supports blinks, a slight head movement, and small changes in expression. A waist-up frame gives breathing room for posture, hands, and clothing. A wider environmental portrait can carry more atmosphere, but it also asks the model to keep more objects coherent.

File rules still matter. MiniMax's current documentation accepts common image formats including JPG, PNG, WEBP, HEIC, and HEIF for first or last frames. It documents a maximum of 30 MB per image, dimensions from 256 to 5760 pixels, and aspect ratios between 2:5 and 5:2. The live site can enforce its own current limits, so its uploader is the final check.

Do not upload a portrait simply because it is available online. Use an image you own, license, or have permission to process. Get the subject's consent when personal data or likeness rights apply. If the finished video could be mistaken for authentic footage, disclose that it is synthetic where platform rules or local law require it.

Compose for the movement you want

The image should contain room for the action. If the prompt asks for a head turn but the face is already pressed against the edge of the frame, the model must invent missing space. If the subject's hands will move, include the entire hands. If wind should catch the hair or coat, make those edges visible.

The window portrait below is a useful starting composition because the face, torso, hands, curtain, and plant are all readable. That does not mean every element should move. Pick one subject action and one environmental reaction. For example, the subject shifts their gaze while the curtain moves slightly. Asking the hands, camera, plant, curtain, lighting, and face to change together adds risk without necessarily adding meaning.

For vertical social video, compose the source vertically rather than forcing a late crop. Leave a little space over the head and around moving fabric. Check what will remain visible inside platform overlays, especially at the top and bottom of the frame.

Write a prompt as a short performance note

The portrait is already doing most of the visual work, so the prompt should describe what changes. A reliable structure is:

subject action + expression or gaze + environmental motion + camera behavior + pace

Put the identity-critical action first. Use concrete verbs. "She slowly turns her eyes toward the window" gives the model a clearer task than "make this cinematic and emotional." Mood words can help, but they should support the action instead of replacing it.

Keep the camera instruction modest. A locked camera or gentle push-in usually gives the face more stability than a large orbit. Avoid prompt clauses that fight the source image. A profile portrait should not instantly face the opposite direction. A seated subject should not stand unless there is enough frame and time to build the movement.

You can also say what must remain stable in plain language: consistent facial features, unchanged wardrobe, continuous background geometry. Negative phrasing is not a guarantee, but it makes the requested boundaries explicit.

Prompt for a subtle close portrait

The woman takes a quiet breath and blinks once, then shifts her gaze slightly toward the window. A few loose strands of hair move in a soft indoor draft. Her facial features, hairstyle, cream knit top, and the apartment background remain consistent. Locked camera, natural pace, restrained movement.

Prompt for a medium environmental portrait

The man looks out of the window, then lowers his gaze slightly as his hands relax. The sheer curtain and plant leaves move gently in the same light breeze. Keep his facial features, dark overshirt, hand anatomy, and room geometry consistent. Very slow camera push-in, realistic timing.

Prompt for a vertical street portrait

The subject turns their attention from the street toward the camera with a small, calm expression change. Short hair and the edge of the wool coat react subtly to the evening breeze. Reflections shimmer softly on the wet pavement. Preserve the face, clothing, and street layout. Vertical portrait framing, stable camera, unhurried motion.

Prompt for a creator opener

The creator settles into frame, gives a brief confident smile, and makes one small hand gesture near the chest. The background remains still and uncluttered. Keep facial identity, fingers, wardrobe, and lighting consistent. Chest-up framing, fixed camera, clean natural motion.

These are starting prompts, not measured presets. Change one variable at a time. If the first clip has a stable face but too little motion, increase the action slightly. If it has lively movement but a drifting face, reduce the head turn or camera move before rewriting everything.

Generate the first clip in the current interface

Open the image-to-video workspace and select MiniMax H3 if it is not already active. Add the portrait in the image input area. Use additional references only when they serve a clear purpose, because every new source adds another instruction for the model to reconcile.

Paste a short motion prompt. Match the aspect ratio to the source composition whenever possible. The public site currently offers MiniMax H3 duration choices from 5 to 15 seconds and 2K output, while MiniMax's API documentation describes H3 generation from 4 to 15 seconds at 768P or 2K. The interface you are using is authoritative for the job in front of you.

For an identity-sensitive first pass, choose the shortest duration that can contain the action. A small gaze change rarely needs 15 seconds. Short clips are easier to diagnose, easier to edit, and less likely to accumulate facial drift. Select resolution based on the destination and available credits. The pricing page shows the current site plans; this guide does not repeat price numbers that may change.

Review the inputs, then generate. Video generation is asynchronous, so a submitted job can take time to finish. Do not interpret a progress state as a completed result, and do not close the page until you know how the current interface preserves jobs.

For a broader walkthrough of the product controls, see How to Use MiniMax H3.

Review the face before the spectacle

Watch the result once at normal speed without scrubbing. Then inspect it again with attention to five zones: eyes, mouth, jawline, hairline, and hands. A clip can feel impressive at first glance while one eye shape changes halfway through or the teeth appear during a closed-mouth expression.

Use this review order:

1. Does the subject still look like the source portrait at the beginning, middle, and end?
2. Do both eyes track the same action without changing shape or color?
3. Does the mouth behave naturally for the requested expression?
4. Does the jaw connect cleanly to the neck through the motion?
5. Do hair, ears, jewelry, and clothing stay attached to the same person?
6. If hands appear, do finger count and position remain plausible?
7. Does the background keep its geometry while the camera moves?
8. Does the motion start and stop naturally enough to edit?

The best take is not always the take with the most motion. For portrait work, a quiet clip with stable identity is often more useful than a dramatic clip that cannot survive a close crop.

Fix common portrait failures

When the face changes, reduce motion before adding more preservation language. Large head rotations reveal parts of the face that the source does not show. The model has to invent them. Ask for an eye shift or slight chin movement instead, or start from a more frontal image.

When the subject looks frozen, add one observable action rather than a stronger mood adjective. Replace "more alive" with "takes a quiet breath and blinks once." If that works, add a small environmental movement in the next version.

When the background swims, remove the camera move or choose a cleaner source. Busy rails, text, patterned walls, and repeated windows expose geometric drift quickly. A locked camera with subject motion can still feel cinematic when light and depth are strong.

When hands deform, crop them out or begin with a source where both hands are fully visible. Keep the gesture small and away from the face. Do not ask for a hand to enter from outside the source frame unless the action is important enough to justify the added uncertainty.

When the portrait becomes glossy or overdramatic, remove stacked style words. One clear lighting instruction and one pace instruction are usually enough. The source image should carry the visual style. The prompt should carry the performance.

When a long clip drifts late, shorten it and build the sequence from several shots. MiniMax H3 can create longer short-form clips, but available duration is not a requirement to use every second. Editability is the practical goal.

Build a sequence instead of one overloaded shot

A portrait video becomes more useful when you plan three small clips rather than one complicated clip. Start with a close shot that establishes the face. Follow with a medium shot that introduces a gesture or environment. End with a detail, profile, or wider view that gives the editor a clean exit.

Keep wardrobe, lighting direction, and color treatment consistent across source images. Reuse a clear reference when the workflow supports it, but expect to review each shot independently. Reference inputs guide the generation; they do not remove the need for continuity checks.

In the edit, trim unstable opening or closing frames. Cut on a blink, gaze shift, or change in posture. Add voice, music, captions, and exact lip synchronization only after the visuals are approved. This order prevents you from spending time on audio timing for a shot that will later be rejected for identity drift.

For more control over wording, use the examples in the MiniMax H3 prompt guide and adapt them to one portrait action at a time. The MiniMax H3 release overview explains the broader multimodal model context.

A practical portrait-video checklist

Before generating:

• Confirm that you own or have permission to use the portrait.
• Choose a source with a clear face, readable pose, and manageable background.
• Match the crop to the final platform.
• Decide on one subject action and, at most, one environmental reaction.
• Keep camera movement small for the first pass.
• Check the live interface for current duration, resolution, file, and credit requirements.

Before publishing:

• Compare the face at the start, middle, and end.
• Check eyes, mouth, jaw, hairline, ears, accessories, and hands.
• Watch the background for sliding lines or changing objects.
• Check that the motion can be trimmed cleanly.
• Confirm that music, voice, captions, and platform crops do not hide defects.
• Label synthetic media where required and avoid misleading uses of a person's likeness.

Create your first portrait clip

The most dependable starting point is intentionally small: one clear portrait, one believable action, one short clip. Let the source image define the person and visual style. Let the prompt define the performance. Review identity before resolution, camera flair, or editing polish.

When you are ready, open the MiniMax H3 image-to-video generator, upload a portrait you have the right to use, and begin with the shortest motion prompt in this guide. A restrained first success gives you something solid to build on.

Sources and methodology

Product capabilities and input limits were checked on August 12, 2026 against the official MiniMax H3 release and MiniMax video generation guide. Rights and consent guidance was checked against MiniMax's current terms and privacy materials. Site controls were checked in the current minimaxh3.tv generator.