MiniMax H3 prompt guide
MiniMax H3 Prompt Guide
Five copyable prompts for MiniMax H3, one for each way the model is driven: a product ad, image-to-video, first and last frame, a video motion reference, and native dialogue with sound effects. Each one names the mode it runs in, the inputs it expects, and what to check in the result.
Availability pending in Imgo
MiniMax H3 is not yet in Imgo's live model catalog, so these prompts are for copying: run them through MiniMax's API or your own H3-Base deployment, or adapt them for a video model that is live in Imgo today.
What you are prompting
Two open checkpoints
H3-Base ships as FL2VA (text, first frame, last frame, or first and last frame to video with audio) and Ref2VA (reference images, video and audio to video with audio). Pick the checkpoint before you write the prompt.
768p locally, 2K through the API
The open weights render 768p. 2K comes from H3-Regenerate-2K, which is not open-sourced, so the full 2K workflow runs through MiniMax's API.
Sound is part of the prompt
Every clip comes back with 32 kHz stereo audio. Write the dialogue, sound effects and ambience into the prompt, or the model will choose them for you.
4–15 seconds, 21:9 to 9:16
Clips run from 4 to 15 seconds at 24 fps, in aspect ratios that include 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16.
Ref2VA reference limits
Stay inside these limits when you assemble a reference-mode request.
| Input | Limit |
|---|---|
| Images | Up to 9 |
| Video clips | Up to 3; each 2–15 s; 15 s in total |
| Audio clips | Up to 3; each 2–15 s; 15 s in total |
| All files | 12 at most across images, video and audio |
First and last frames belong to FL2VA. A single run is conditioned on frames or on references, never on both.
Copy and adapt
Product ad from packshots
A 10-second 16:9 commercial for [product]. Use the reference images as the exact product: keep its shape, colors, logo position and label text unchanged in every shot. Open on a slow push-in across [surface] in [lighting]; then a hand lifts the product and turns it to show [key feature]. End on a centered hero frame with clean negative space on the right. Audio: a [music style] music bed and a soft [product sound] as the product is lifted; no voiceover. No on-screen text and no extra products.
- Mode
- Ref2VA · reference images
- Use it when
- A real product has to stay exactly right while the camera, the set and the sound are invented around it.
- Inputs
- Two to four photos of the real product from different angles, with the label legible.
- Check
- Compare the label, logo and color against the packshots, and listen for the product sound landing on the lift.
Bring a still image to life
Animate the first-frame image over 6 seconds. Keep [subject], the composition and the color grade exactly as they are. [subject] [action] while the camera [camera move]. Background motion: [ambient motion]. Audio: the natural ambience of [setting] and [sound effect] in sync with the movement; no music and no dialogue. Do not add people, objects or text.
- Mode
- FL2VA · first frame
- Use it when
- The image is already approved and needs to move without being redrawn.
- Inputs
- One high-resolution image in the target aspect ratio; it becomes frame one.
- Check
- Compare the opening frame with your image, then watch hands, faces and edges for drift as the motion builds.
First frame to last frame
Create an 8-second transition from the first-frame image to the last-frame image as one continuous shot. Keep [subject] identical throughout. Motion: [transition action], while the camera [camera move]. Timing: hold the opening for one second, move steadily, and settle on the final frame for the last second. Audio: [transition sound] building under the motion, landing on a soft [final sound] as the last frame settles.
- Mode
- FL2VA · first and last frame
- Use it when
- The start and end states are fixed — a before and after, a reveal, a transformation — and only the path between them is open.
- Inputs
- Two images with the same aspect ratio, framing and lighting; the closer they match, the cleaner the path between them.
- Check
- Confirm the clip really lands on the last frame and that the subject keeps its identity halfway through.
Borrow motion from a reference clip
Use the reference video only for its camera movement and timing: [camera move description]. Do not copy its people, set or colors. Apply that motion to [subject] from the reference image, in [setting] with [time of day] light, and keep the appearance of [subject] identical to the reference image. Match the length of the reference clip. Audio: [ambient sound] and [motion sound] timed to the movement; no music.
- Mode
- Ref2VA · reference video and images
- Use it when
- You like a camera move or a piece of choreography and want it on a different subject.
- Inputs
- One 2–15 s reference clip that shows the motion, plus one to three images of the new subject.
- Check
- Play the reference and the result side by side: the move should match, and nothing from the reference's set or cast should leak through.
Dialogue scene with sound effects
A 12-second 16:9 scene in [setting]. [character A], [description A], and [character B], [description B], sit across a table. Medium two-shot, then a slow push-in on [character A]. Dialogue in [language]: [character A] says, "[line 1]" and [character B] answers, "[line 2]". Sound effects: [sound effect 1] before the first line and [sound effect 2] under the reply. Ambience: [ambience]. Lips move in sync with each line; no music and no subtitles.
- Mode
- FL2VA · text only
- Use it when
- The clip has to speak: lines, voices and effects written into the shot instead of added in the edit.
- Inputs
- Text only. Keep each line short enough to say comfortably in the time the shot allows.
- Check
- Check lip sync line by line, that each voice stays with its speaker, and that the effects land where the prompt places them.
Practical tips
- Choose the checkpoint first: frames go to FL2VA and references go to Ref2VA. One run never takes both.
- Count files before you upload: at most 9 images, 3 video clips and 3 audio clips, and 12 files in total.
- Trim reference clips to the part that matters. Each video or audio clip must run 2–15 s, and each type is capped at 15 s combined, so three 5-second clips fill the budget.
- Say what each reference is for — product, motion, voice or palette — so the model does not copy everything in it.
- Write sound like a shot list: who speaks, which effects, which ambience, and whether there is music. Asking for silence is an instruction too.
- Spoken lines are most stable in the 11 languages MiniMax lists: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish.
- Draft at 768p and spend 2K only on the takes worth keeping; 2K output runs through MiniMax's API.
Frequently asked questions
- What is the best prompt structure for MiniMax H3?
- Subject and action, then camera, then setting and light, then audio, then exclusions. In Ref2VA mode, open by saying what each reference is for. MiniMax also publishes base and reference prompt-writing guides on the H3 model card.
- Can I generate MiniMax H3 videos in Imgo?
- Not yet. H3 is not in Imgo's live model catalog, so its availability is shown as pending. Copy the prompts here to run them through MiniMax or your own deployment, or adapt them for a video model that is live in Imgo.
- Why does my local H3 render stop at 768p?
- The open H3-Base checkpoints render 768p. 2K comes from H3-Regenerate-2K, a second pass that MiniMax has not open-sourced, so the full 2K workflow runs through MiniMax's API.
- Can I use the open weights commercially?
- Self-hosting falls under the MiniMax H3 Community License. It does not cover the EU, the UK, the Republic of Korea or the USA (organizations there can apply to MiniMax for a license); products earning more than US$20 million a year need written authorization from MiniMax via [email protected]; commercial products must display "MiniMax H3" in their user interface; and Hong Kong law governs the agreement. Access through MiniMax's API follows MiniMax's API terms instead. This is a summary, not legal advice — read the license before you ship.
- Do the prompts work in other languages?
- Yes. Describe the shot in any language. For spoken lines, H3 is most stable in the 11 dialogue languages MiniMax lists; other languages work to varying degrees.
Keep going
Sources
Reviewed on October 7, 2026. Model limits and license terms can change; the linked sources are authoritative.