MiniMax H3 Prompt Tips: Structure, Limits and What a Render Costs

The cheapest way to iterate on MiniMax H3 is to draft at 768P and regenerate the keeper to 2K — it costs exactly the same as rendering at 2K from the start, but buys you a review pass in between. At $0.08 per second for 768P output and $0.05 per second to regenerate that clip to 2K, a 15-second shot runs $1.20 + $0.75 = $1.95. Going straight to 2K at $0.13 per second is $1.95 for the same 15 seconds. Identical cost, one extra chance to catch a bad take.

That arithmetic shapes almost every prompting decision below, so it is worth stating before the craft advice.

What H3 actually reads

According to the MiniMax video generation guide, H3 accepts three entry modes, and the prompt does different work in each:

Mode	Inputs	What the prompt controls
Text-to-Video	Prompt only	Everything — subject, camera, pacing, style
Image-to-Video	Prompt + first and/or last frame	The motion between fixed endpoints
Reference Generation	Prompt + reference images, video, or audio	How the references are assigned and combined

The third mode is where most people underuse the prompt. References do not carry their own labels. If you hand H3 four images, the prompt is the only place that says which one is the character and which one is the product.

The pattern in MiniMax's own examples

Every worked example in MiniMax's H3 feature highlights follows the same shape, and it is worth copying literally:

Lead with duration and aspect ratio. The examples open with fragments like 15s, 9:16 before any description.
Assign each reference a role by number. Image 1: mood, setting, film grain. Image 2: character. Image 3: bag. Image 4: ending logo. Numbered role assignment is the single most repeated device in the official set.
Then tell the story. The examples break to Core story: or Story: and describe the action in one continuous sentence rather than a shot list.
Name the camera behaviour explicitly. Subtle handheld shake, first-person, eye-level, handheld gameplay camera, smooth vertical scroll — camera language is stated, never implied.

The ordering matters because the format constraints are non-negotiable and the story is not. Put the fixed facts first.

Constraints that will silently cost you a render

These come from the model spec table, and several of them are the kind that fail a job after you have already paid for the queue:

Duration is an integer. H3 accepts 4 to 15 seconds; H3 Max accepts 5 to 15. Not 7.5.
Resolution differs by model. H3 outputs 768P or 2K. H3 Max outputs 480P or 768P — it has no 2K tier, so the draft-then-upgrade trick above does not apply to it.
Reference images cap at 9, videos at 3 clips, audio at 3 clips.
Mixed input is capped at 12 files in total, which is lower than the sum of the individual caps. A nine-image board plus three reference clips is already at the ceiling with no room for an audio reference.
Per-clip reference duration is 2 to 15 seconds and total reference video must stay under 15 seconds.
Dimensions must sit in 256–5760 px with an aspect ratio between 2:5 and 5:2, and files cap at 50 MB for video, 30 MB for image, 15 MB for audio.

The 12-file ceiling is the one that changes prompting strategy. When you cannot bring every reference, the prompt has to carry the roles you dropped — which is exactly why the numbered-role device exists.

Cost math for planning a shoot

Output billing is per second of generated video, so the prompt's duration field is also its budget field. Rates below are from MiniMax's pay-as-you-go pricing:

Job	Rate	15-second clip
H3 output at 2K	$0.13 / second	$1.95
H3 output at 768P	$0.08 / second	$1.20
Regenerate 768P to 2K	$0.05 / second	$0.75

Input material is cheaper than most people expect: audio references are free, the first five input images are free, and additional images are $0.04 each. Input video is billed by its duration against the output resolution, which means a long reference clip can quietly cost more than the shot you are making.

The practical consequence: a board of nine reference images costs $0.16 — four paid images beyond the free five. Spending that to avoid one failed 2K render is an easy trade.

Where the free input allowance changes the workflow

Because audio references cost nothing and the first five images are free, the cheap experiment is to over-specify references and under-specify prose, rather than the reverse. If a scene keeps drifting, adding a mood reference image is free; re-rendering is not.

This inverts the habit most people bring from text-to-image tools, where prompt engineering is the free variable and everything else is fixed.

FAQ

What is the correct prompt structure for MiniMax H3? Open with duration and aspect ratio, assign each reference a numbered role such as Image 1: character, Image 2: setting, then describe the action as one continuous story, and state camera behaviour explicitly. This is the structure MiniMax uses in every worked example in its own H3 highlights gallery.

Is it cheaper to render H3 at 768P and upgrade to 2K? It costs the same. A 15-second clip is $1.20 at 768P plus $0.75 to regenerate to 2K, which equals the $1.95 of rendering at 2K directly. The reason to draft at 768P is the review pass, not the price.

How many reference files can MiniMax H3 take? Up to 9 images, 3 video clips, or 3 audio clips — but mixed input is capped at 12 files in total, which is lower than the individual limits added together.

Does MiniMax H3 Max support 2K output? No. H3 Max outputs 480P or 768P only, and its duration floor is 5 seconds rather than H3's 4. If you need 2K, you need H3.

Why does my H3 job reject a 7.5-second duration? Duration must be an integer. H3 accepts 4 to 15 seconds and H3 Max accepts 5 to 15, in whole seconds only.

For the wider Chinese video and multimodal model landscape, see the MiniMax provider page.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing. Last verified: 2026-09-15.