Featured Example
Disaster scene prompt
MiniMax H3 disaster-style clip — useful for ads, shorts, and dramatic B-roll.
MiniMax H3 Reference-to-Video
Control Video with Image, Video & Audio References
Give each reference a clear job. Use images to define identity, appearance, products, or scenes; videos to guide motion, performance, camera behavior, or timing; and audio to reference voice, rhythm, or sound. MiniMax H3 combines those references into a new audiovisual video.
If you're searching for "MiniMax H3 reference video," this is H3's Reference-to-Video workflow: multiple image, video, and audio references can each contribute different information to the final generation.
Last verified: August 9, 202610 Free Credits Daily 路 Up to 30 Free Credits
First frame0/1
Last frame (optional)0/1
Reference images
0/5Reference videos
0/3MP4 / MOV / MKV, up to 3 files, 50MB each
Reference audio
0/3MP3 / WAV / AAC / FLAC, up to 3 files, 15MB each
0/2500
2K
4s
Est. cost: ~48 creditsBalance: — credits
Checking your account…
OutputVideo
History
Sample videos
Hover to preview on desktop; tap to load and play on mobile.
Hover to preview, click to expand
Hover to preview, click to expand
Hover to preview, click to expand
Hover to preview, click to expand
Hover to preview, click to expand
Hover to preview, click to expand
Prompt examples
Featured MiniMax H3 examples — hover to preview, tap to view the prompt.
Featured Example
Naval battle action
Fast-paced action with storm and combat effects from a text-only prompt.
Featured Example
Anime character scene
Original anime-style motion from text and reference imagery.
Featured Example
Sci-fi mecha action
Complex camera moves and mecha combat described in natural language.
Featured Example
Cinematic transformation
Dark fantasy scene with film lighting and camera motion in the prompt.
Featured Example
Bullet-time racing
Slow-motion rain and orbiting camera around a race car.
What Is MiniMax H3 Reference-to-Video?
MiniMax H3 Reference-to-Video generates a new video from a text prompt plus any combination of reference images, reference videos, and reference audio. Each reference can provide a different part of the target result, such as character identity, clothing, style, body motion, camera movement, voice, or scene context.
The most important part is not how many references you upload. It is whether each reference has a clear role.
More references do not automatically create more control. Clearer reference roles create more control.
What Is a MiniMax H3 Reference Video?
A MiniMax H3 reference video is a source clip used to guide part of a newly generated video. It can provide body movement, acting performance, camera movement, cuts, rhythm, or other temporal information.
A reference video does not automatically mean that H3 is editing the original clip. If the source video only provides motion, camera behavior, or timing, it is being used as reference guidance rather than as footage that must be directly edited.
MiniMax H3 Reference-to-Video vs Video-to-Video
| Workflow | Use It When | Core Input |
|---|---|---|
| Reference-to-Video | Multiple images, videos, and audio files each provide different information for a new generation | Mixed multimodal references |
| Video-to-Video | One existing source video is the center of the task and you want to transform, edit, or transfer from it | Existing source video |
| Image-to-Video | A supplied image should become the opening or ending frame | First / last frame image |
Use Video-to-Video when one source clip is the main thing you want to transform or edit. Use Reference-to-Video when multiple reference assets need to work together, such as an image for identity, a video for motion, and audio for voice.
MiniMax H3 Reference Input Limits
| Reference Type | Current Limit |
|---|---|
| Reference images | Up to 9 |
| Reference videos | Up to 3 |
| Reference audio clips | Up to 3 |
| Reference video duration | 2-5 seconds each |
| Total reference-video duration | Up to 15 seconds |
| Reference audio duration | 2-5 seconds each |
| Total standalone audio duration | Up to 15 seconds |
| Reference video formats | MP4, MOV |
| Reference image formats | JPG, JPEG, PNG, WEBP, HEIC, HEIF |
| Output resolution | 768P or 2K |
| Generated duration | 4-15 seconds |
These are maximum input limits, not a recommendation to use every available slot.
Reference audio needs visual context. Standalone audio references should be paired with at least one visual reference such as an image or video.
MiniMax3.org Reference Assignment Map
MiniMax3.org practical framework
Before generating, give each reference one clear job. Avoid asking several assets to control the same property unless they are intentionally reinforcing one another.
| Reference Asset | Best Job |
|---|---|
| Character portrait | Face and identity |
| Full-body character image | Identity, body appearance, clothing |
| Outfit image | Clothing and accessories |
| Product image | Shape, materials, colors, branding |
| Scene image | Environment, layout, composition |
| Style image | Visual style or art direction |
| Motion video | Body movement |
| Performance video | Acting, gesture, expression timing |
| Camera video | Tracking, orbit, push-in, cuts, shot rhythm |
| Storyboard image | Composition and shot planning |
| Voice audio | Voice timbre and delivery |
| Music reference | Rhythm or music style |
| Sound reference | Sound texture or audible continuity |
One reference, one primary job.
Multiple roles are possible, but role clarity should come before reference quantity.
A Reference File Is Not the Same as a Subject
A reference file is the source asset. A Subject is the reusable visual content extracted from one or more references. One image can define multiple subjects, and one subject can combine information from several assets.
Picture 1: A woman wearing a black jacket beside a motorcycle.
Subject 1: The woman's facial identity.
Subject 2: The black jacket and silver accessories.
Subject 3: The motorcycle.
Subject 1: The woman's appearance comes from Picture 1, while her walking motion comes from Video 1.
Picture 1 does not automatically equal Subject 1.
How H3 Reference Labels Work
| Label | What It Represents |
|---|---|
| Subject N | Reusable visible content such as a person, object, scene, clothing, style, action, or pose |
| Picture N | A specific reference image used as a frame, keyframe, storyboard, or composition anchor |
| Video N | A reference video used for editing, continuation, motion, camera, cuts, rhythm, or temporal structure |
| Audio N | Audio that is copied or referenced for voice, music, rhythm, dialogue, or sound |
A Video label represents the whole-video relationship. If a character from that video is reused as visible content, that character should still be treated as a Subject rather than assuming Video 1 is the character itself.
How Many References Should You Actually Use?
MiniMax3.org practical guidance
H3 supports many references, but most production tasks do not need the maximum number. Start with the smallest reference set that fully defines the result.
Simple Character Scene
1 identity image + optional 1 motion video
Character + Outfit
1 identity image + 1 clothing image + optional motion video
Character + Motion + Voice
1 identity image + 1 motion video + 1 audio reference
Product Commercial
1 product image + optional scene/style image + 1 camera or motion video + optional audio
For a complex multi-reference scene, only add another reference when it contributes information that the current reference set does not already provide.
Use the smallest reference set that completely specifies the task.
What Happens When H3 References Conflict?
Reference conflicts happen when multiple assets give H3 incompatible instructions about the same part of the target video.
| Conflict Type | Example |
|---|---|
| Identity conflict | Two images define different faces for the same character |
| Clothing conflict | One image shows a black coat while another defines a white dress |
| Motion conflict | Two videos provide incompatible body movements |
| Camera conflict | Prompt asks for a static camera while Video 1 provides aggressive handheld movement |
| Scene conflict | Several references define incompatible environments |
| Voice conflict | Multiple audio references try to define the same speaker differently |
More references can create less control when their roles conflict.
MiniMax3.org Reference Conflict Check
- Which asset defines identity?
- Which asset defines appearance or clothing?
- Which asset defines movement?
- Which asset defines camera behavior?
- Which audio defines each speaker or sound role?
If two references answer the same question differently, resolve the conflict in the prompt before generating.
How to Write a MiniMax H3 Reference-to-Video Prompt
Start by defining what each reference represents. Then describe how those roles should appear in the target video.
Reference - Role - Relationship - Target Shot
4 Copyable MiniMax H3 Reference Video Prompts
Simple Character + Motion Prompt
StarterPicture 1 defines the main character's facial identity, hairstyle, and clothing. Video 1 provides the walking motion and slow tracking-camera behavior. Generate a new cinematic night-street scene that preserves the character from Picture 1 while following the movement and camera rhythm from Video 1. The character walks through a rain-soaked neon street while the camera tracks backward at the same pace. Preserve the character's identity throughout the shot.
Character + Outfit + Motion Prompt
StarterPicture 1 defines the character's facial identity and hairstyle. Picture 2 defines the black leather jacket, silver necklace, and clothing details. Video 1 provides the body movement and camera timing. Generate a new video that combines the identity from Picture 1, the outfit from Picture 2, and the movement from Video 1. Preserve all three roles consistently while changing the environment to a rainy train platform at night.
Character + Motion + Voice Prompt
StarterPicture 1 defines the main character's identity and appearance. Video 1 provides the walking motion, gesture timing, and camera movement. Audio 1 provides the target voice timbre and measured delivery. Generate a new cinematic street scene that preserves the character identity from Picture 1, follows the movement and camera behavior from Video 1, and uses Audio 1 only as the voice reference for the character's new dialogue.
Product + Camera Reference Prompt
StarterPicture 1 defines the product's exact shape, materials, color, and branding. Video 1 provides the slow camera arc and push-in timing. Generate a new studio product video that preserves the product design from Picture 1 while following the camera movement from Video 1. Keep the background clean and dark, with controlled highlights moving naturally across the product surface.
Advanced H3 Full-Reference Prompt Structure
For advanced reference workflows, H3's full-reference structure separates the prompt into six sections.
| Section | Purpose |
|---|---|
| subject_definitions | Define subjects and reference roles |
| summary | Summarize the task and main relationships |
| retention_analysis | Explain what is preserved, transferred, copied, or referenced |
| detailed_description | Describe the target video shot by shot |
| overall_soundscape | Describe ambience and physical sound |
| non_diegetic_music | Describe audience-only background music |
Most users do not need to manually write the full structure for every generation. The practical goal is to make the same relationships clear: what each reference represents and how it should affect the target video.
MiniMax H3 Prompt GuideHow Strongly Should a Reference Be Preserved?
| Intent | Practical Meaning |
|---|---|
| Fully preserve | Keep the referenced property as closely as possible |
| Partially preserve | Keep the reference but allow controlled changes |
| Transfer attribute | Move a referenced characteristic to another subject |
| Weak reference | Use only broad similarity such as style, atmosphere, or composition |
Do not ask every reference to be fully preserved. Some references only need to guide style, movement, camera, or mood.
Should You Use an Image, Video, or Audio Reference?
| Need | Better Starting Reference |
|---|---|
| Face identity | Image |
| Outfit detail | Image |
| Product shape | Image |
| Scene appearance | Image |
| Body motion | Video |
| Acting performance | Video |
| Camera movement | Video |
| Cuts or shot rhythm | Video |
| Voice timbre | Audio |
| Dialogue delivery | Audio |
| Music rhythm or sound texture | Audio |
MiniMax H3 Reference-to-Video Output
| Feature | Visible Page Claim |
|---|---|
| Native audiovisual generation | H3 can generate video with native stereo audio |
| Resolution | Use 768P for iteration or 2K for final high-detail output |
| Duration | Generate short clips from 4 to 15 seconds |
| Reference images | Use images for identity, appearance, product, scene, and style control |
| Reference video | Use video for motion, performance, camera, cuts, and timing |
| Reference audio | Use audio for voice, rhythm, music style, and sound continuity |
Common MiniMax H3 Reference Video Mistakes
- Uploading many references without assigning roles.
- Expecting every uploaded asset to be fully preserved.
- Using a reference video when you actually need to edit the original clip.
- Letting two references define the same character differently.
- Asking audio alone to create the full Reference-to-Video scene.
- Combining complex motion, complex camera, and several identity constraints in a very short clip.
MiniMax H3 Reference Video Troubleshooting
| Problem | Try This |
|---|---|
| Identity drifts | Assign one image as the identity source and reduce conflicting appearance instructions |
| Motion is ignored | State exactly what Video 1 contributes, such as walking motion or camera push-in |
| Camera feels wrong | Separate subject motion from camera behavior in the prompt |
| Voice is inconsistent | Use one audio reference per speaker or voice role |
| Too many details change | Reduce the number of simultaneous reference requirements |
What Can You Make with MiniMax H3 Reference Video?
Consistent Character Video
Use an image for identity while a motion video defines how the character moves.
Motion Transfer
Use a reference video to transfer body movement, performance, or camera behavior into a new character or scene.
Product Commercial
Combine a product image with a camera-reference video to create controlled commercial movement while preserving product appearance.
Character + Voice
Combine a character image with motion and voice references so appearance, performance, and audio each come from a clearly assigned source.
MiniMax H3 Reference Video FAQ
What is MiniMax H3 Reference-to-Video?
MiniMax H3 Reference-to-Video generates a new video from a text prompt plus reference images, videos, and audio. Each reference can provide a different part of the result, such as character identity, style, motion, camera movement, or voice.
What is a MiniMax H3 reference video?
A reference video is a source clip used to guide motion, acting, camera movement, cuts, rhythm, timing, or other temporal properties in a newly generated H3 video.
How many reference images can MiniMax H3 use?
Current MiniMax H3 Reference-to-Video supports up to 9 reference images.
How many reference videos can H3 use?
H3 supports up to 3 reference videos. Each reference video can be 2-5 seconds long, and their total reference-video duration can be up to 15 seconds.
How many audio references can H3 use?
H3 supports up to 3 standalone reference audio clips. Each clip can be 2-5 seconds long, with up to 15 seconds of total standalone reference-audio duration.
Is a reference image the same as a Subject?
No. The reference image is the source asset. A Subject is the reusable visible content defined from one or more references. One reference image can define multiple subjects, and one subject can combine information from several assets.
Should I use Reference-to-Video or Video-to-Video?
Use Reference-to-Video when multiple image, video, and audio references each provide different information for a new generation. Use Video-to-Video when one existing source video is the center of the task and you primarily want to transform, edit, or transfer from that clip.
Can I use a video only for motion?
Yes. A reference video can provide movement, performance, camera behavior, cuts, rhythm, or timing without requiring the original footage itself to be directly edited.
Can I use one image for identity and another for clothing?
Yes. Giving different references separate jobs can make the intended relationship clearer.
Can MiniMax H3 use image, video, and audio together?
Yes. H3 Reference-to-Video supports mixed multimodal references, allowing images, videos, and audio to contribute different information to the same target generation.
Does MiniMax H3 Reference-to-Video generate audio?
Yes. MiniMax H3 is an audiovisual generation model with native stereo audio.
Can H3 Reference-to-Video generate 2K?
Yes. Current H3 Reference-to-Video supports 768P and 2K output.
How should I prompt MiniMax H3 Reference-to-Video?
Define what each reference represents, assign each reference a clear job, and then describe how those roles should appear together in the target video.
Create with MiniMax H3 References
Start with the smallest set of references that fully defines your target video. Assign each image, video, and audio file a clear role, then tell H3 how those roles should work together in the final shot.
10 free credits per daily check-in 路 Up to 30 total