How to Actually Prompt MiniMax H3 (It's Not Like Other AI Video Models)
MiniMax H3's hosted platform quietly rewrites your prompts into a structured format before generation. Here's what that format actually is, and how to write it yourself.
MiniMax H3 generates video and audio from a single prompt, both signals come from the same denoising pass, which is why dialogue lands on the exact syllable instead of getting stitched together in post. But that joint generation only works well if the prompt itself is structured the way the model was actually trained to read, and most people prompting H3 casually are leaving real capability on the table without realizing it.
Why Casual Prompts Underperform
MiniMax's own official model card confirms something worth knowing directly: when you use H3 through MiniMax's hosted product, your casual prompt passes through a preprocessing system called H3-Context-IR, which rewrites it into the model's expected structured format before generation happens. MiniMax's documentation states plainly that this system "is critical to the quality of the final output," and recommends either using it directly or replicating its structure yourself.
That preprocessor isn't open-sourced, and it isn't present when H3 is accessed through platforms running the base model directly. In that case, whatever structure you write is exactly what the model sees, no automatic cleanup happening behind the scenes. This guide's structural breakdown comes largely from deAPI, an AI API aggregator that published a detailed reverse-engineered prompting guide; worth knowing deAPI has a commercial interest in this framing, since the same post promotes their own not-yet-launched H3 access. The underlying technical claims, though, trace back to MiniMax's own documentation and are worth taking seriously regardless of who's explaining them.
The Three Required Fields
Every H3 prompt is built around three named sections, written in this exact order, not optional labels, structure the model was specifically trained on.
integrated_multimodal_description is the main body: visual style, composition, camera movement, actions, shot changes, and any dialogue or sound the characters themselves can hear. It's written in playback order, starting with [Shot 1] establishing visual style and composition. Style keywords the model responds well to include cinematic, live-action, 2D-animated, 3D CG, claymation, and watercolor.
overall_soundscape covers ambient sound, physical action noise, and non-verbal human sound across the whole clip, wind, footsteps, fabric rustling. Dialogue and music don't belong here; they have their own fields.
non_diegetic_music describes score the audience hears but characters don't. This field specifically rejects abstract mood language, "tense emotional music" doesn't work. Naming actual instruments, tempo, and how the music changes over time does. Use "N/A" when a scene should have no score at all.
Prompt Length Matters More Than You'd Expect
H3's text encoder (Qwen3-VL-32B) is built to consume dense, detailed instruction, not brief description. Aim for 350-450 words in the main description field for complex scenes, and 150-250 for simpler single-shot clips. A 40-word prompt leaves the model filling in gaps on its own, which is exactly where casual prompting tends to underperform against the structured format.
Shot Structure and Timing
The first shot never gets a timestamp. Every subsequent shot needs one, in MM:SS.mmm format, strictly increasing. The practical rule worth internalizing: decide total clip duration first, then place your cuts, then write the content, working in the opposite order tends to produce timestamps that fall outside the actual clip length.
Budget roughly one cut per 3 seconds, with each shot needing at least 3 seconds to establish anything meaningful. Camera movement gets written as natural action inside a sentence ("the camera pushes in with small amplitude at slow speed") rather than as trailing tags, and H3 supports a specific set of named movements: zoom, push/pull, pan, truck, tilt, pedestal, arc, tracking, static, shake, POV, and roll.
The Dialogue System Is Genuinely Strict
H3 generates spoken dialogue with matching lip movement across 11 languages. Anyone who speaks gets a stable speaker ID, (S1), (S2), assigned in order of first appearance and kept consistent across every shot. Spoken words go inside dialogue tags with a language marker, separate from the speaker description and delivery notes around them.
Word budget matters directly: natural speech runs about 2.5 words per second, so a 10-second clip realistically fits 20-25 spoken words if anything else needs to happen in the scene. Overloading dialogue causes H3 to either rush delivery or cut it off. One detail worth remembering specifically: voiceover narration requires explicitly stating the on-screen character's lips stay closed, skip that line and H3 will animate their mouth to match narration it was never meant to lip-sync.
The Core Principle: Everything Must Be Observable
This rule underlies the entire format. Abstract emotional or mood language doesn't translate into anything the model can act on. "She feels abandoned" doesn't work; "she lowers her gaze and her shoulders drop" does. "Melancholic atmosphere" doesn't work; "rain streaks the window, grey light fills the room" does. The same logic applies to the music field: naming instruments and describing how they change produces better results than mood adjectives ever will.
Specs Worth Knowing
- Duration: 4-15 seconds
- Frame rate: 24 fps (fixed)
- Resolution: 768p native (short edge); 2K available only through MiniMax's own hosted module
- Audio: 32 kHz stereo, generated jointly with video
- Dialogue languages: 11 (Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish)
- Max prompt length: 7,000 characters
- Negative prompts: not supported
The Signal in the Noise
The structured prompt format isn't bureaucratic overhead, it's the actual language H3 was trained to read, and the gap between a casual prompt and a properly structured one appears to be real and measurable based on the documented before/after comparisons. This connects directly to the broader pattern worth understanding across AI video models generally: the platforms and community resources that reverse-engineer and publish a model's actual expected syntax, rather than treating prompting as trial and error, are producing genuinely more useful guidance than most generic "prompt engineering tips" content. Worth testing this structure directly against your own casual prompts before assuming H3's output ceiling is lower than it actually is.
Have you tested H3 with a fully structured prompt yet? Curious how much of a difference it actually made for you, drop it in the comments.
