What Is MiniMax H4 Video-to-Video?
MiniMax H4 video-to-video describes workflows where an existing video is central to the result. H3 can use that clip for body motion, acting performance, camera movement, cuts, rhythm, timing, editing context, or continuation.
The source video’s role determines the workflow. If the clip only teaches H3 how something should move, it is motion or temporal reference. If the original footage itself must remain recognizable while selected elements change, it is video editing.
What Do You Want H3 to Do with the Source Video?
| Your Goal | Best Interpretation | What H3 Should Preserve |
|---|---|---|
| Transfer body movement to a new character | Motion transfer / reference generation | Motion and timing |
| Copy camera orbit, tracking, or handheld behavior | Camera reference | Camera path and shot rhythm |
| Replace the person but keep the action | Character replacement | Motion, camera, timing |
| Change clothing, props, lighting, or background in the existing clip | Video editing | Most non-target footage |
| Modify dialogue or performance while retaining the scene | Video editing | Scene, camera, identity as requested |
| Generate what happens after the source clip | Video continuation | Ending state and temporal continuity |
Do not start by asking “Is this V2V?” Start by deciding what the source video is supposed to contribute to the result.
Motion Transfer vs Video Editing: What’s the Difference?
Motion transfer uses the source video as behavioral or temporal reference. The final video can keep the movement, performance, camera path, or timing while changing the character, environment, or visual style.
Video editing treats the existing footage itself as the object being modified. The target is usually to preserve most of the source clip while changing only the elements requested in the prompt.
When Is Video Continuation the Better Workflow?
Use continuation when the new generation should begin from the state reached at the end of the source clip and extend the action forward in time.
A continuation prompt should describe what happens next rather than repeatedly describing the earlier part of the source. Focus on the next action, the camera transition, scene continuity, sound continuity, and the target ending state.
What Should H3 Take From Your Source Video?
MiniMaxH4.org practical framework
| Source Information | Best Role |
|---|---|
| Body movement | Motion reference |
| Gesture timing | Performance reference |
| Facial performance | Performance reference |
| Camera tracking | Camera reference |
| Orbit / push / pull movement | Camera reference |
| Cuts | Temporal structure |
| Shot duration | Timing reference |
| Rhythm / pacing | Temporal reference |
| Existing scene and subjects | Video editing source |
| Final state of the clip | Continuation starting point |
| Existing soundtrack | Audio reuse when required |
| Character appearance alone | Prefer a clear image reference when available |
A source video does not need to control every property of the result. Stronger prompts tell H3 exactly which parts of the source matter.
The MiniMaxH4.org Preserve / Transfer / Change Method
MiniMaxH4.org practical framework
Before writing a video-to-video prompt, separate the request into three decisions.
1. Preserve — What Must Stay?
List the source properties that should remain stable.
Examples: body motion, camera path, timing, shot order, environment, facial identity, original audio, object placement.
2. Transfer — What Should Be Reused?
Define which source behavior should move into the new result.
Examples: movement, acting performance, gesture timing, camera behavior, pacing, voice delivery, editing rhythm.
3. Change — What Must Become Different?
State exactly what the final video should change.
Examples: character, clothing, environment, lighting, props, dialogue, visual style, product.
MiniMax H4 Video-to-Video Prompt Examples
Motion Transfer Prompt
StarterVideo 1 provides the full body movement, acting rhythm, and camera timing. Preserve the movement sequence and overall shot timing. Picture 1 defines the replacement character’s facial identity, hairstyle, and clothing. Transfer the performance from Video 1 to the character from Picture 1 while generating a new cinematic night-street environment. Keep the body mechanics natural and maintain consistent character identity throughout the shot.
Camera Transfer Prompt
StarterUse Video 1 only as reference for camera movement, shot timing, and pacing. Preserve the tracking path, camera speed, and timing, but do not preserve the original subject or environment. Generate a new scene featuring a woman in a black coat walking through a rainy neon train station. Follow the camera behavior from Video 1 while allowing the new character and location to be generated naturally.
Targeted Video Editing Prompt
StarterEdit Video 1. Preserve the original subject identity, body movement, camera framing, timing, background, and overall composition. Change only the subject’s blue jacket to a deep red leather jacket. Keep all other visible elements, motion, lighting relationships, and scene timing as close to the source as possible.
Character Replacement Prompt
StarterVideo 1 provides the body motion, camera movement, timing, and scene structure. Picture 1 defines the replacement character’s facial identity, hairstyle, body appearance, and clothing. Replace the original person with the character from Picture 1 while preserving the source movement sequence, camera path, and shot timing. Maintain consistent identity throughout the full clip.
Is Your Source Video Good for Motion Transfer?
MiniMaxH4.org practical checklist
| Check | Better Source | Higher-Risk Source |
|---|---|---|
| Subject visibility | Full moving body is clearly visible | Key body parts are frequently cropped |
| Motion clarity | Actions and contact points are distinct | Heavy blur or ambiguous movement |
| Occlusion | Hands and feet remain visible | Limbs disappear behind objects |
| Camera behavior | Camera path is easy to identify | Erratic movement with no clear pattern |
| Shot structure | One coherent action or shot sequence | Many unrelated cuts |
| Subject separation | Character is visually separated from background | Subject blends into background |
| Action completion | Movement has a clear start and finish | Clip cuts off mid-action |
How to Replace a Character Without Losing the Source Motion
For character replacement, separate appearance from movement. Let the video define how the character moves and let a clear image reference define who the replacement character is.
Video 1 = motion + camera + timing
Picture 1 = identity + appearance
Prompt = relationship between the two
If the source character’s appearance is not important, do not ask H3 to preserve it at the same time that you are asking for a replacement character.
Can MiniMax H4 Transfer Camera Movement?
Yes. A source video can provide camera behavior such as tracking, push-in, pull-back, orbiting motion, handheld movement, cuts, or shot rhythm without requiring the original character or environment to remain.
In the prompt, explicitly state that the source video’s camera path and timing should be preserved while the subject or scene changes.
MiniMaxH4.org Edit Locality
MiniMaxH4.org benchmark method
A good targeted video edit should change the requested element while preserving unrelated parts of the source.
Edit Locality = Preserved non-target requirements ÷ Total non-target requirements
Example
Requested edit: Blue jacket → red jacket
Check:
✓ Face unchanged
✓ Body movement unchanged
✓ Camera unchanged
✓ Background unchanged
✓ Timing unchanged
✓ Audio unchanged
Do not publish a numerical H3 score until MiniMaxH4.org has run and documented a first-party test.
MiniMaxH4.org methodology only — no first-party H3 score published yet.
MiniMaxH4.org Motion Transfer Fidelity
MiniMaxH4.org benchmark method
Instead of judging motion transfer only by overall visual quality, break the source action into observable checkpoints and test whether the generated result reproduces them.
Example
Source motion checkpoints:
- Step left
- Raise right arm
- Turn 180 degrees
- Crouch
- Finish in a fixed pose
Motion Checkpoint Completion = Correctly reproduced checkpoints ÷ Total checkpoints
This metric evaluates instruction and motion adherence rather than visual polish alone.
MiniMaxH4.org methodology only — no first-party H3 score published yet.
Video-to-Video vs Reference-to-Video vs Image-to-Video
| Workflow | Choose It When |
|---|---|
| Video-to-Video | One existing source clip is the center of the transformation |
| Reference-to-Video | Several image, video, and audio references each have different jobs |
| Image-to-Video | A supplied image must anchor the opening or ending frame |
Use Video-to-Video when the question is “What should happen to this clip?”
Use Reference-to-Video when the question is “How should these different references work together?”
Use Image-to-Video when the question is “How should this still image move?”
MiniMax H4 Video Output
MiniMax H4 is a multimodal audiovisual video model that understands text, image, video, and audio context. Current H3 generation supports 4–15 second output at 768P or 2K with native stereo sound.
The source video’s role can vary across reference generation, editing, motion transfer, camera reference, and continuation-style workflows.
7 Common MiniMax H4 Video-to-Video Mistakes
- Not telling H3 what the source video is for.
- Asking H3 to preserve and replace the same property at the same time.
- Treating motion transfer and video editing as the same workflow.
- Using the source video to define precise character identity when a clear image reference would be better.
- Asking for major changes to character, camera, scene, timing, and dialogue in one uncontrolled prompt.
- Using a source clip where the important motion is hidden by blur, cropping, or occlusion.
- Judging the result only by visual quality instead of checking whether the requested motion or edit was preserved.
MiniMax H4 Video-to-Video Troubleshooting
| Problem | Try This |
|---|---|
| Motion is not preserved | Explicitly assign Video 1 as the motion reference |
| Character identity drifts | Add one clear image reference for identity |
| Camera changes too much | State that the source camera path and timing must remain |
| Too much of the original video changes | Rewrite the request as a targeted edit and list what must stay |
| Output copies the source too closely | Clarify which properties should change and which are only weak references |
| Motion looks chaotic | Use a cleaner source clip with fewer cuts and clearer body visibility |
| Edit changes unrelated objects | Explicitly list the non-target elements that must remain unchanged |
| Character replacement changes timing | Lock source movement, camera, and temporal structure |
| Voice changes unexpectedly | Explicitly define whether source audio should be preserved, replaced, or ignored |
What Can You Do with MiniMax H4 Video-to-Video?
Motion Transfer
Transfer body movement, acting performance, or gesture timing from an existing clip to a newly generated character or scene.
Character Replacement
Use the source clip for movement and camera behavior while a reference image defines the replacement character.
Targeted Video Editing
Modify selected elements of existing footage while preserving the parts that should remain unchanged.
Camera Transfer
Reuse tracking, orbiting, handheld movement, push-ins, cuts, or shot rhythm while generating a different subject or environment.
MiniMax H4 Video-to-Video FAQ
Does MiniMax H4 support video-to-video?
Yes. MiniMax describes H3 as supporting generalized multimodal reference and editing and specifically highlights V2V motion transfer. A source video can provide motion, camera behavior, timing, editing context, or continuation context.
What is MiniMax H4 video-to-video?
MiniMax H4 video-to-video is a broad user-facing term for workflows where an existing video is central to the result. The source can guide motion or camera behavior, act as the footage being edited, or provide temporal context for continuation.
What is MiniMax H4 motion transfer?
Motion transfer uses a source video to guide movement, acting performance, gesture timing, camera behavior, or temporal structure in a new generation.
Is motion transfer the same as video editing?
No. Motion transfer uses the source mainly as behavioral or temporal reference. Video editing modifies the source footage itself while preserving the parts that should remain unchanged.
Can MiniMax H4 replace a character in a video?
H3 can combine source-video motion or camera information with image references that define another character’s identity and appearance. The prompt should explicitly assign those roles.
Can MiniMax H4 copy camera movement from a video?
Yes. A reference video can provide camera movement, cuts, rhythm, and temporal structure while allowing the subject and environment to change.
How should I prompt MiniMax H4 video-to-video?
First state what the source video is for. Then separate the instruction into what must be preserved, what should be transferred, and what should change.
Should I use Video-to-Video or Reference-to-Video?
Use Video-to-Video when one source clip is the center of the transformation. Use Reference-to-Video when several images, videos, and audio references each need different control roles.
Should I use Video-to-Video or Image-to-Video?
Use Video-to-Video when you already have footage that should influence the result. Use Image-to-Video when a supplied still image should anchor the generated video’s opening or ending frame.
Does MiniMax H4 generate audio?
Yes. MiniMax H4 is an audiovisual generation model with native stereo sound.
How long can MiniMax H4 videos be?
Current MiniMax H4 generation supports output durations from 4 to 15 seconds.
Does MiniMax H4 support 2K?
Yes. Current MiniMax H4 generation supports 768P and 2K output.
Transform Your Source Video with MiniMax H4
Start with one source clip and decide what H3 should preserve, transfer, or change. Use the video for motion, camera behavior, editing context, or temporal structure, then describe the target result clearly.
10 free credits per daily check-in · Up to 30 total