Chrome ExtensionMiniMax H3 Video Prompt Generator
Write prompts for MiniMax H3, the model that replaced Hailuo and generates video with synchronized audio in a single pass. This generator follows MiniMax's published prompt-writing guides: the three-section format, the natural-language camera grammar that replaced Hailuo's bracket commands, speaker and dialogue tags for lip-sync in eleven languages, timed multi-shot cuts, and explicit roles for every reference asset.
Describe the main visual scene — subjects, environment, mood, and key visual details
Quick Palettes
Generated Prompt
Fill in the form and click "Generate" to create an optimized MiniMax H3 video prompt.
Tip: Describe the motion and temporal progression of your scene. Think in terms of "what happens over time" rather than a static description.
MiniMax H3 Tips
- • Hailuo bracket commands like [Push in] are gone. H3 wants camera motion as natural English inside the sentence
- • Camera direction is type plus amplitude plus speed: "pushes in with small amplitude at slow speed"
- • Say "holds a static shot" out loud when you want a locked frame, because H3 drifts by default
- • Write audio in three places: diegetic sound in the shot, ambience in the soundscape, score as non-diegetic music
- • Cut only when new information arrives. If just the distance changes, move the camera instead
- • Give every reference asset an explicit job rather than uploading images and hoping
- • There is no seed and no negative prompt on the hosted API, but prose negation works well
The bracket commands are gone
If you learned Hailuo 02, unlearn this part. H3 does not take stacked camera labels. Camera motion belongs inside the sentence as natural English, built from three pieces.
Hailuo 02 style, no longer correct
A woman looks out of a train window. [Push in] [Truck left] [Static shot]
H3 style
The camera trucks right with small amplitude at slow speed as she lifts her gaze from the folded letter toward the passing city lights.
| Motion type | zoom in, zoom out, push in, pull out, pan left, pan right, truck left, truck right, tilt up, tilt down, pedestal up, pedestal down, arc shot, tracking shot, static shot, shake slightly, shake strongly, POV, roll clockwise, roll counterclockwise |
|---|---|
| Amplitude | with small amplitude, with large amplitude. Omit for medium |
| Speed | at slow speed, at fast speed. Omit for normal |
| Put together | The camera pushes in with small amplitude at slow speed toward the folded letter in her hands |
The official three-section format
Every base-mode prompt MiniMax publishes uses these three labelled sections. Splitting sound into diegetic, ambient and score is what keeps the mix under control.
integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium-wide shot frames a baker opening the shutters of a small street bakery before sunrise. The camera pushes in with small amplitude at slow speed as the baker (S1) places a fresh loaf on the counter and says: "First batch of the morning." [Shot 2] At 00:05.000, the camera cuts to a close-up of steam rising from the sliced bread. overall_soundscape: Wooden shutters scrape open over a quiet street as trays clink softly inside the bakery. The doorbell rings once, followed by light footsteps and the crisp sound of bread being sliced. non_diegetic_music: A soft acoustic-guitar pattern at a moderate tempo, joined by sparse upright-bass notes and a gentle fade at the end.
MiniMax H3 specs and configs
| Model | 33B dense single-stream Omni Transformer. API model ID MiniMax-H3 |
|---|---|
| Released | July 31, 2026. Open weights August 2, 2026 |
| Resolution | 768P or 2K, where 2K is a 1440px short edge. No 1080p tier, no 4K |
| Duration | 4 to 15 whole seconds, default 5. fal enforces a minimum of 5 |
| Frame rate | 24fps fixed. Audio at 32 kHz stereo |
| Aspect ratios | adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16. Text-to-video needs a concrete ratio |
| Modes | Text-to-video, first frame, last frame, first plus last, and reference-to-video. Image and reference modes are mutually exclusive |
| References | Up to 9 images, 3 videos and 3 audio files, capped at 12 files total. Audio cannot be the only reference |
| Audio languages | Eleven with stable support: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish |
| Prompt limit | 7,000 characters per text item |
| Parameters | model, content, resolution, duration, ratio, callback_url. No seed, no negative prompt, no guidance scale, no step count |
| Pricing | $0.13 per second at 2K, $0.08 per second at 768P on MiniMax pay-as-you-go |
| Licence | MiniMax H3 Community License. Weights exclude the EU, UK, South Korea and USA. The API remains globally available |
How to Use the Prompt Generator
Pick the mode and the shot structure
Choose text-to-video, a keyframe mode, or reference-to-video, then decide between a single shot and a timed multi-shot sequence. If you are using references, give each one an explicit job in the reference roles field.
Direct the picture and the sound
Describe the subject and action, pick one camera motion with its amplitude and speed, and write your audio. Keep diegetic sound with the action, ambience in the soundscape, and any score as non-diegetic music.
Generate, then draft at 768P
Copy the prompt into the MiniMax platform, Hailuo AI, fal or ComfyUI. Render at 768P first since it costs roughly 60% of 2K, then regenerate your pick at 2K once the timing and performance look right.
Frequently Asked Questions
What is MiniMax H3?
MiniMax H3 is MiniMax's current flagship video model, released on July 31, 2026 as the successor to the Hailuo family. It is a 33B dense single-stream Omni Transformer that generates video and synchronized audio in the same pass. The API model ID is MiniMax-H3. You will see it called Hailuo 3.0 or Hailuo 03 on some sites, which comes from the internal API naming, but MiniMax H3 is the official name. Weights were opened on Hugging Face on August 2, 2026.
Did the Hailuo bracket camera commands survive into H3?
No, and this is the biggest change to how you write prompts. Hailuo 02 used bracket tokens like [Push in], [Truck left] and [Pan right], stacked up to three at a time. H3 does not work that way. MiniMax's own prompt-writing guides state camera motion should be written as a natural English action inside the shot rather than stacked as labels at the end. In H3 the square brackets are reserved for shot numbers, as in [Shot 1]. One page of MiniMax's platform guide still mentions the old bracket style, but every official prompt-writing document and every published example uses natural language, so treat that line as leftover copy.
How do I write camera movement for H3?
Three parts in one natural sentence: motion type, then amplitude, then speed. For example, "the camera pushes in with small amplitude at slow speed toward the folded letter in her hands." The supported motions are zoom in and out, push in and pull out, pan left and right, truck left and right, tilt up and down, pedestal up and down, arc shot, tracking shot, static shot, shake slightly or strongly, POV, and roll clockwise or counterclockwise. Omit the amplitude and speed phrases when you want medium and normal. One continuous move follows far better than several abrupt ones, and if you want a locked-off frame you have to say "holds a static shot" explicitly, because the model drifts otherwise.
What is the official prompt format?
Three labelled sections. First integrated_multimodal_description, which carries style, composition, action, camera and dialogue, with shots marked as [Shot 1], [Shot 2] and so on. Then overall_soundscape, one to four sentences covering ambience, physical action sounds and non-verbal human sounds. Then non_diegetic_music, one to three sentences on instrumentation, tempo, rhythm and dynamics, or simply N/A when you want none. Reference-to-video uses a longer six-section format that adds subject definitions, a summary tagged with the task type, a retention analysis, and a detailed description of roughly 350 to 500 words.
Does MiniMax H3 generate audio?
Yes, natively and in the same pass as the picture, at 32 kHz stereo. Dialogue, foley, room tone and score are modelled jointly rather than dubbed on afterwards, which is why lip-sync holds up. Speakers are tagged (S1), (S2) or (S1,S2), and dialogue is wrapped as a d tag carrying the language, so a line reads like [English] followed by the text. There is a tag for lines that cross a cut and another for speech that gets truncated. Stable support covers eleven languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish.
How do multi-shot prompts work?
Multi-shot is native, so several coherent cuts come out of a single generation. Number the shots and give each cut a timecode, as in "[Shot 2] At 00:05.000, the camera cuts to a close-up." Writing a timed shot list is the main defence against pacing drift, where a clip slowly turns into a slideshow. The useful discipline is to cut only when new information arrives: if all that changes is distance or a slight angle, move the camera instead of cutting.
What resolutions and durations does H3 support?
Two resolution tiers, 768P and 2K, with no 1080p tier and no 4K. The 2K tier means a 1440 pixel short edge, so 2560 by 1440 at 16:9, and it is produced by feeding the 768p result and the original context back through the model rather than by a separate upscaler. Duration runs 4 to 15 seconds as whole numbers with a default of 5, though fal enforces a minimum of 5. Frame rate is fixed at 24. Aspect ratios are adaptive, 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, with the note that text-to-video needs a concrete ratio and cannot use adaptive, while image-to-video ignores the setting and follows your input image.
Can I use a seed or a negative prompt?
Not on the hosted API. The request body is limited to model, content, resolution, duration, ratio and an optional callback URL, and that is the whole surface. There is no seed unless you self-host, no negative prompt parameter, and no guidance scale or step count because the released checkpoints are CFG-distilled. Negation in prose does work well though, and MiniMax's own examples use it, with lines like "no soft dissolves or fluid morphs" or "do not introduce garbled characters." There is also no prompt optimizer flag; instead MiniMax sells a separate H3-Context-IR endpoint that rewrites your prompt into the structured format and returns it without generating video.
How do reference-to-video prompts work?
You can attach up to 9 reference images, 3 reference videos and 3 reference audio files, capped at 12 files in total. The rule that matters most is to give every asset an explicit job. "Use Image 1 for mood and location, Image 2 for the talent, Image 3 for the bag, Image 4 for the closing brand mark" beats uploading four images with a general description. Each reference also carries a retention marker saying whether it is fully preserved, partially preserved, an attribute transfer or a weak reference. Note that image-to-video and reference-to-video are mutually exclusive, so you cannot mix first or last frame images with reference assets in one request. Audio cannot be the only reference either, it needs an image or video alongside it.
How long can a MiniMax H3 prompt be?
Up to 7,000 characters per text item, which is generous enough for a full timed shot list plus detailed sound design. That said, length is not the goal. The reference format suggests roughly 350 to 500 words for the detailed description, and the published examples are tightly written rather than padded.
Where does H3 rank against other video models?
Well, and on independent evaluation rather than self-reported numbers. On the Artificial Analysis Video Arena, which uses blind human preference, H3 is first overall in video editing at 1132 Elo, second in text-to-video with audio at 1238 behind Gemini Omni Flash, and third in image-to-video with audio. It is the top open-weights model in every one of those arenas. Worth flagging: Black Forest Labs published charts claiming FLUX 3 beats H3, but those come from BFL's own evaluation with no published methodology or sample size and sit on a different scale from the public arena, so the two sets of numbers are not comparable. MiniMax itself published no benchmark numbers at all.
What does MiniMax H3 cost, and where can I run it?
On MiniMax's own pay-as-you-go pricing it is $0.13 per second at 2K and $0.08 per second at 768P, with regeneration from 768P to 2K at $0.05 per second. Your first five input images are free, then $0.04 each, and input audio is free. Third-party hosts charge more, with fal at roughly double the official rate. You can run it through the MiniMax platform API, Hailuo AI on the web, fal, Runware, Segmind, Krea, WaveSpeed, ComfyUI, or locally through SGLang and vLLM. It is not on OpenRouter, and we could not verify a working Replicate listing.
Are the open weights usable everywhere?
No, and this is widely misreported. The MiniMax H3 Community License defines an applicable territory that excludes the European Union, the United Kingdom, South Korea and the United States, and that exclusion covers use, distribution and outputs of the open weights. Redistribution requires displaying "Powered by MiniMax H3", commercial products must show the name prominently, and organisations above $20M annual revenue need separate written authorisation. The important correction is that this restricts the weights only. MiniMax states the API stays globally available, and has publicly pushed back on claims that H3 cannot legally be used in those regions. Also note the closed sub-models: only the 768p base is open, while the context compiler and the 2K regeneration model are not.
Is this tool free to use?
Yes. You get 3 free prompt generations per day with no signup required. For unlimited access, sign up for a Promptslove membership which includes all AI tools and 20,000+ premium prompts.
More video prompt generators
Want Unlimited AI Prompt Generation?
Get unlimited access to all AI tools, 20,000+ premium prompts, courses, and resources designed to maximize your creative output.
MiniMax, MiniMax H3 and Hailuo are trademarks of their respective owners. Promptslove is not affiliated with or endorsed by MiniMax. Model specifications, pricing and licence terms described here reflect published documentation at the time of writing and may change.