How long can a MiniMax H3 clip be?

MiniMax's model card lists an output duration of 4 to 15 seconds at 24 FPS, with 32 kHz stereo audio. The shorter side is 768 pixels by default, and 2K output is produced by a separate module, H3-Regenerate-2K, which feeds the 768p result plus the original context back through the model rather than running a conventional upscaler. Supported aspect ratios include 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. Two prompt-side consequences follow: every shot timestamp you write must fall inside the duration and increase strictly, and a clip in that range still only holds about one beat per shot, so a second action means a second shot.

Does MiniMax H3 generate audio, and how do you prompt it?

Yes. H3 generates 32 kHz stereo audio jointly with the picture, and the prompt format splits that audio across three places instead of one. Dialogue, singing, and any music or sound a character can actually hear go in integrated_multimodal_description at the moment they happen. Ambient sound, physical action sounds, and non-verbal human sounds such as breathing or laughter go in overall_soundscape, in one to four sentences. Score that only the audience hears goes in non_diegetic_music, in one to three sentences describing instrumentation, tempo, rhythm, and dynamics rather than mood. Write N/A in non_diegetic_music when you want no score; that is H3's equivalent of appending "no music" to a Veo prompt.

How do you write an image-to-video prompt for MiniMax H3?

Image-to-video is the I2VA task, and it starts with a fixed line: "For the target video, at 0.00 seconds into the target video, (from [Shot 1]) is fully referenced." That line comes first, then one blank line, then the three core fields. Shot 1 should re-establish what the image already fixes, meaning subject, clothing, colors, key objects, and spatial relationships, so the model holds them, and then describe what happens next. MiniMax's recommended shape is first-frame anchor, action onset, continuous development, then result or reaction. Derive the style from the reference image instead of inventing a new one, or the render drifts away from the frame you approved.

Is MiniMax H3 the same thing as the Hailuo app?

Hailuo is MiniMax's consumer video brand and app; H3 is the model behind it, also sold as MiniMax-H3 on the platform API and carried by partners under names including Hailuo 03 and hailuo-3. The prompt-craft difference is what sits in front of the model. Hosted surfaces run H3-Context-IR, a preprocessing system that reads your free-form text and reference files and rewrites them into the structured three-field representation the base model actually consumes. Write that structure yourself and you replace the rewriter's inferences with your own decisions. Specs, pricing, and the open-weights situation are covered in our MiniMax H3 explainer.

Can I run MiniMax H3 locally?

Partly. MiniMax publishes two BF16 checkpoints on Hugging Face, MiniMax-H3 Base FL2VA for text and first/last-frame tasks and MiniMax-H3 Base Ref2VA for reference tasks, under the MiniMax H3 Community License, with deployment paths for SGLang, vLLM, diffusers, and ComfyUI. Two modules are not in that release. H3-Context-IR, the hosted prompt-preprocessing system, is not open-sourced, so nothing rewrites your prompt locally and you write the structured format yourself. H3-Regenerate-2K, which produces the 2K output, is also not yet open-sourced, so a local deployment gives you 768p. MiniMax provides an API for reproducing the full 2K workflow.

How is prompting H3 different from Veo 3.1 or Seedance?

The three dialects are not interchangeable, and pasting one into another fails silently because every model reads unknown syntax as plain text. Veo 3.1 takes a single plain-language paragraph with no token syntax, labels audio inline with SFX: and Ambient noise:, and rewards brevity. Seedance 2.0 takes a six-slot description and tags references numerically, as [Image 1] or @Image1 depending on the surface. H3 takes named fields, timestamped shots, camera motion quantified by amplitude and speed, and stable speaker IDs, and it is the only one of the three that separates diegetic from non-diegetic sound as a formal field. What carries over unchanged: one beat per shot, and dialogue as an exact quoted line rather than a topic.

What do the , , and tags do?

They are special tokens in H3's tokenizer, not decoration. wraps spoken content: inside it goes only a language tag and the exact words, preserved verbatim including punctuation and never translated, while everything identifying or performative, meaning the speaker ID, voice description, action, and delivery, stays outside the tag. marks the connecting points on both sides of a cut when one line of dialogue or lyrics carries across it, and it should be paired with an explicit continuity phrase such as "continues seamlessly across the cut". marks speech that the end of the video truncates. For voiceover, use the exact phrase "says in an off-screen voiceover" and state right after the block that the character's lips stay closed.

Guides

MiniMax H3 Prompts: The Official Format, With Examples

MiniMax H3 prompts are three labeled fields, not a sentence. The official format explained, with five copy-ready prompts for text, image, and keyframe video.

By Aditya BawankuleAugust 5, 202613 min readUpdated August 11, 2026

Frame from MiniMax's official H3 text-to-video sample: a starship captain on the bridge as the fleet's hyperdrive flash floods the observation window behind her.

MiniMax H3 prompts are not sentences. They are a three-field document, and MiniMax publishes the exact shape it wants: integrated_multimodal_description for everything you can see and hear on the timeline, overall_soundscape for ambience, and non_diegetic_music for the score. Write those three fields properly and H3 does what you asked. Paste in a one-line prompt written for another model and you are handing the picture, the dialogue, and the mix to a guess.

This guide is a working walkthrough of MiniMax's own video prompt writing guide, shipped in the H3 repository on Hugging Face, plus five complete prompts you can copy and edit. For what the model is, what it costs, and where the weights stand, read our MiniMax H3 explainer first; this post is only about writing for it.

Why H3 wants a structured prompt at all

The format exists because of how the full system is assembled. MiniMax's model card describes three modules: H3-Context-IR, a hosted preprocessing system that reads your free-form text, images, video, and audio and rewrites them into a Context Intermediate Representation; H3-Base, which generates 768p video and audio from that representation; and H3-Regenerate-2K, which feeds the 768p result plus the original context back through the model to produce 2K.

Context-IR is not part of the open-weights release, and MiniMax "strongly recommends" you either call its hosted Context-IR API or follow the prompting guidance to build your own preprocessing system. The format below is therefore the intermediate representation the base model is trained to read, not a stylistic preference. Run H3 locally through SGLang, vLLM, diffusers, or ComfyUI and nothing rewrites your prompt; use the hosted API and writing it yourself replaces the rewriter's inferences with your decisions.

One expectation to reset first. Advice for Veo and Seedance tells you to stay short, because those models over-direct when you stack instructions. H3 goes the other way: the reproducible examples in MiniMax's own model card run from roughly 300 to 700 words per request, and the field structure is what keeps that length from turning into contradiction. Density is the point, as long as every sentence describes something visible or audible.

The four task types

The base guide covers four tasks, each beginning with a different first line. T2VA has no image-alignment instruction; the other three open with a literal alignment sentence, then one blank line, then the three fields.

Task	What it does	Opening line
T2VA	Builds the whole audiovisual timeline from text	None. Start at `integrated_multimodal_description`
I2VA	Your image is the first frame; the video develops forward from it	`For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.`
FL2VA	Two images anchor the opening and the ending; you supply the path between them	An alignment sentence naming both pictures and their timecodes
L2VA	One image anchors only the final frame; the model infers a plausible earlier state	An alignment sentence naming the picture and the ending timecode

In the FL2VA and L2VA lines, the shot index is the actual final shot and the timecode is the effective duration written to exactly two decimal places, so an eight-second clip aligns its last picture to the 8.00 mark. The examples further down show both lines in full.

Each task also has a recommended internal shape, and it is what stops the model stalling or rushing. I2VA runs first-frame anchor, action onset, continuous development, result. FL2VA runs first-frame state, observable intermediate changes, narrowing differences, last-frame state, and prefers a single shot so the model can interpolate continuously. L2VA runs a plausible preceding state, an explicit transition path, then the last-frame landing.

Field one: integrated_multimodal_description

This is the body of the prompt, and one rule governs it: every detail must correspond to something visible or audible. Style, composition, subject appearance and position, props, actions, cuts, spoken language, and synchronized diegetic sound live here. Intentions, backstory, and mood adjectives do not.

Shot 1 carries no timestamp and opens with the overall style. The guide names Cinematic, live-action, 2D-animated, 3D CG, claymation, watercolor, and vintage film as common choices, and for keyframe tasks you derive the style from the reference image rather than inventing one: [Shot 1] Live-action, cinematic, a medium-wide shot frames...

Later shots open with a strictly increasing cut time inside the duration: [Shot 2] At 00:03.500, the camera cuts to... The accepted phrasings for an ordinary cut are "the camera cuts to", "the shot cuts to", "the shot transitions to", "the shot changes to", and "the shot switches to"; cross-dissolve, fade, and wipe are reserved for when you ask for them explicitly. The editorial rule attached is the useful part: a cut has to introduce new information about the subject, space, state, viewpoint, or time. If all you want is a different distance or angle, move the camera instead of cutting.

Camera motion needs three dimensions

H3 wants motion type, amplitude, and speed. Add amplitude and speed only when they matter: medium amplitude and normal speed are the assumed defaults and are usually left out.

Motion type: Zoom In and Zoom Out (focal length changes, body still), Push In and Pull Out (camera moves forward or back), Pan Left and Pan Right, Tilt Up and Tilt Down, Truck Left and Truck Right, Pedestal Up and Pedestal Down, Arc Shot, Tracking Shot, Static Shot, Shake Slightly and Shake Strongly, Roll Clockwise and Roll Counterclockwise, and POV.
Amplitude: with small amplitude or with large amplitude.
Speed: at slow speed or at fast speed.

The taxonomy separates things most prompt guides blur. Zoom and push are different instructions, and so are pan and truck, so "zoom in slowly" and "push in with small amplitude at slow speed" produce different shots. Write the motion as an action inside the sentence rather than a label stapled to the end: The camera pushes in with small amplitude at slow speed toward the folded letter in her hands. The single-move rule holds here as on every other video model. One camera instruction per shot; a second move means a second shot.

Speakers, dialogue, and voiceover

Anyone who speaks, sings, or produces an off-screen human voice gets a stable ID: (S1), (S2), and a compound (S1,S2) when already-numbered speakers vocalize together. The ID follows the character across shots, and characters who never make a sound get no ID.

The split around the <d> tag is the part people get wrong. Everything identifying, physical, or performative goes outside the tag: character type, age, gender, whether they are on screen, pitch, timbre, speaking rate, accent, and the action performed while speaking. Inside goes only a language tag and the exact words, verbatim including punctuation, never translated.

The young woman with a quiet, breathy voice (S1) says: <d>[English] I get off at the next station.</d>
The two children (S1,S2) shout together, <d>[English] Wait for us!</d>

Voiceover has a required phrase. Use says in an off-screen voiceover, and immediately after the <d> block state that the on-screen character's lips stay closed, or you get a character mouthing narration they are not supposed to be speaking. Two more tags handle edges: <scenetrans> goes at the connecting points on both sides when one line of dialogue or lyrics crosses a cut, paired with an explicit continuity phrase such as "continues seamlessly across the cut", and <cutoff> marks speech the end of the video truncates.

On-screen text

Any banner, sign, label, subtitle, or neon text that is actually visible goes in English double quotation marks, with the original wording and punctuation preserved and untranslated. A Chinese shop sign stays in Chinese inside the quotes: A red neon sign reading "营业中" glows above the doorway. The quotes signal characters to render rather than a sign to describe.

Fields two and three: splitting the audio

H3 generates 32 kHz stereo audio jointly with the picture, and the format splits that audio across three places rather than one. Getting the split right is the difference between a clean mix and a clip that scores itself over your dialogue.

overall_soundscape takes one to four sentences in a single paragraph covering ambient sound, physical action sounds, and non-verbal human sounds across the whole video: wind, rain, traffic, footsteps, fabric, impacts, breathing, laughter, panting. Dialogue, singing, and diegetic music are excluded because they already live in the description. Use N/A only when total silence is what you want.

non_diegetic_music takes one to three sentences describing only the score the characters cannot hear. Give instrumentation, tempo, rhythm, and dynamic change, and skip abstract mood words. "Sparse piano notes at a slow tempo, joined by sustained low strings that gradually increase in volume before fading out" is the register. Write N/A when there should be no score, the H3 equivalent of appending "no music" to a Veo prompt.

The trap sits between those two fields. Music a character can hear, from a radio, a phone, or a live instrument, is a diegetic event and belongs in integrated_multimodal_description at the moment it happens. Put a car stereo in the music field and you have asked for a score instead of a car stereo.

That is a lot of structure to hold on a first attempt, and the fastest way to learn it is to write one prompt properly and then break it on purpose: strip the camera amplitude, move the radio into the music field, drop the speaker ID, and watch what each one costs you. H3 is one of the engines in the Dream Pixel Forge video generator, on the same credit balance as Veo 3.1, Seedance 2.0, and Grok Imagine Video, so those runs sit side by side.

MiniMax H3 prompt examples

Five complete prompts in the official format, one per task type plus a second text-to-video case with dialogue across a cut. Change the nouns and keep the shape.

Every one of them was then run through H3, and the clip it produced sits directly under it, so you can check the format against the result rather than taking our word for it. Nothing was regenerated to get a nicer take: each clip is the first render of the prompt as printed. Turn sound on — the audio fields are half of what the format controls, and a muted clip hides the half these prompts work hardest on.

1. UGC product clip (T2VA, vertical)

integrated_multimodal_description: [Shot 1] Live-action, cinematic, a chest-up vertical shot frames a woman in her late twenties with dark shoulder-length hair and a cream knit sweater on a pale linen sofa in a sunlit living room, holding a matte green skincare bottle at chest height. The camera pushes in with small amplitude at slow speed as she turns the bottle so the label faces the lens. The woman with a warm, slightly husky voice and an unhurried speaking rate (S1) looks into the lens and says: <d>[English] Three weeks in, and my skin finally calmed down.</d> She sets the bottle on the oak coffee table and rests her hand beside it.
overall_soundscape: Quiet room tone carries faint traffic through an open window. Fabric shifts as she moves, and the bottle meets the table with a soft knock.
non_diegetic_music: N/A

H3, six seconds, 9:16, generated from the prompt above with no reference image of any kind. Run the audio back through a transcriber and the spoken line comes out as “Three weeks in, and my skin finally calmed down.” — word for word what sits inside the <d> tag, punctuation included. The push-in stays small, and she completes the last beat in the description by setting the bottle down.

2. Two-speaker dialogue across a cut (T2VA)

integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium two-shot frames two colleagues on either side of a counter in a small office break room under flat fluorescent light. The camera holds a static shot as the man in a grey henley with a close-cropped beard and a low, level voice (S1) sets down a coffee pot and says: <d>[English] They moved the deadline again.</d> The woman in a navy blazer with a bright, quick-paced voice (S2) looks up from her laptop and answers, <d>[English] Then we cut the third feature,</d> <scenetrans> [Shot 2] At 00:05.000, the shot cuts to a close-up of the woman as she keeps speaking, and her line <scenetrans> continues seamlessly across the cut: <d>[English] nobody used it anyway.</d> She closes the laptop lid with one hand.
overall_soundscape: A steady fluorescent hum and the low rumble of a refrigerator fill the room. The coffee pot clinks against the counter, a chair creaks, and the laptop lid closes with a soft click.
non_diegetic_music: N/A

Eight seconds, 16:9. Two things to check. The cut arrives on the 00:05.000 mark the prompt names, and it cuts to the close-up of the woman rather than anywhere else. And the sentence the <scenetrans> tag carries across that cut stays unbroken: transcribed, her line runs from three seconds to seven, straight through the edit, with “nobody used it anyway” landing on the far side as one continuous thought.

3. Animating a product photo (I2VA)

For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Live-action, commercial product photography, the matte black espresso machine shown in <Picture 1> stands on the walnut counter exactly as framed, preserving its position, the white ceramic cup beneath the spout, the raking window light from the left, and the dark tiled backsplash behind it. The camera pushes in with small amplitude at slow speed toward the cup as a thin stream of espresso begins to fall from the spout. Pale crema builds into a ring on the surface while a single curl of steam rises and bends toward the window. The stream thins, then stops, leaving the cup two thirds full.
overall_soundscape: A low pump hum runs under the shot, followed by the steady hiss of espresso striking ceramic and a faint metallic tick as the pump cuts out.
non_diegetic_music: A single sustained electric-piano chord at a slow tempo, decreasing in volume through the second half.

Six seconds, 16:9, with the still described in the prompt supplied as Picture 1. The framing, the raking window light, the walnut counter, and the cup under the spout all hold while the stream starts and the steam rises, which is the whole point of the alignment sentence: nothing in the description competes with the frame you already approved. This is also one of only two examples here whose non_diegetic_music field is not N/A, so it is worth listening to next to the three that asked for silence.

4. First and last frame transition (FL2VA, eight seconds)

A grey wool throw folded on the arm of a linen sofa in late-afternoon light, the cushions empty — The two pictures the FL2VA prompt below cites. They have to be the same room, the same framing, and the same light, or the model is being asked to interpolate between two different scenes rather than across one action — so the second was produced by editing the first, not generated separately.

The same sofa in the same light with the grey wool throw shaken open and spread across the seat — The two pictures the FL2VA prompt below cites. They have to be the same room, the same framing, and the same light, or the model is being asked to interpolate between two different scenes rather than across one action — so the second was produced by editing the first, not generated separately.

How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot 1) aligns with the 8.00-second mark of the target video.
integrated_multimodal_description: [Shot 1] Live-action, cinematic, a folded grey wool throw rests on the bare arm of a linen sofa in the position and framing established by Picture 1, late-afternoon light falling across the empty cushions. The camera trucks left with small amplitude at slow speed as two hands lift the throw, shake it open, and lower it across the seat. The fabric settles in stages and the corners are pulled square until the sofa reaches the arrangement, drape, and composition established by Picture 2 at the end of the shot.
overall_soundscape: Soft room tone with faint birdsong through a window. Heavy wool snaps once as it is shaken open, then rustles as it settles into place.
non_diegetic_music: Widely spaced acoustic-guitar notes at a slow tempo, joined by a quiet sustained pad that fades out at the end.

Eight seconds, 16:9. One caveat on how this was run: the video surface it went through takes an opening frame plus ordinary reference images and has no dedicated last-frame slot, so Picture 2 was passed as a reference and the alignment sentence did the rest. It held anyway — the clip opens folded, the throw is lifted and snapped open mid-shot, and it settles into the spread arrangement by the end rather than drifting somewhere else.

5. Landing on a final frame (L2VA, six seconds)

A pale oak kitchen table under low window light with an enamel mug, a squared stack of unopened mail, and a white bowl of cut strawberries — The single picture the L2VA prompt cites, and the only frame in it the model is told to hit. Everything before it — the bare table, the hand, the order the objects arrive in — is the model's to invent.

How the reference pictures align with the target video — <Picture 1> (from [Shot 1]) aligns with the 6.00-second mark of the target video.
integrated_multimodal_description: [Shot 1] Live-action, cinematic, a close shot begins on an empty enamel mug and a loose stack of unopened mail at the edge of the same pale oak kitchen table visible in <Picture 1>, under the same low window light. The camera pushes in with small amplitude at slow speed as a hand enters from the left, sets a white bowl of cut strawberries beside the mug, and squares the mail into a clean stack. The hand withdraws from frame and the objects settle into the exact spacing, angles, lighting, and final composition established by <Picture 1>.
overall_soundscape: Quiet kitchen room tone with a faint refrigerator hum. The bowl sets down on wood and paper slides against paper as the stack is squared.
non_diegetic_music: N/A

Six seconds, 16:9, run the same way as the clip above, with the anchor supplied as an ordinary reference. This is the one that shows what the L2VA shape buys you. It does not open on the picture and hold there. It opens on a loose spill of mail, a hand comes in from the left with the strawberries, the stack gets squared, the hand withdraws — and the last frame is the composition it was told to reach.

Reference mode uses a different format

Everything above comes from the base guide, which covers the four keyframe and text tasks. H3's omni-reference mode, where you feed up to nine images, three video clips, and three audio clips within a twelve-file cap, has its own guide and its own fields: subject_definitions labels each reference as <Subject N>, <Picture N>, <Video N>, or <Audio N>, summary names the task type, and retention_analysis states per reference whether it is fully_preserved, partially_preserved, an attribute_transfer, or a weak_reference. The main body is renamed detailed_description, with the style stated before Shot 1 rather than inside it. For video editing, continuation, or voice matching, that is the guide to read.

How this differs from Veo and Seedance prompting

The three dialects are not portable, and copying one into another fails invisibly because every model reads unknown syntax as plain text. Veo 3.1 takes a plain-language paragraph with no token syntax, labels audio inline with SFX: and Ambient noise:, and rewards brevity. Seedance 2.0 takes a six-slot description and tags references numerically. H3 takes named fields, timestamped shots, quantified camera motion, and speaker IDs, and it is the only one that separates diegetic from non-diegetic sound as a formal field.

Three habits carry over. One beat per shot, because 4 to 15 seconds at 24 FPS holds no more. Dialogue is an exact quoted line rather than a topic, which the <d> tag enforces more strictly than any other model does. And when you already have a still you approved, describe motion and let the frame carry composition, the argument in our guide to image to video AI.

The short version

Pick the task type and write its opening line. Open Shot 1 with the style, timestamp every later shot, give each shot one camera move with amplitude and speed, put verbatim dialogue inside <d> and everything about the speaker outside it, quote on-screen text without translating it, then fill the soundscape and score fields separately, using N/A deliberately rather than leaving a field blank. That is the whole MiniMax H3 prompt guide in a paragraph.

If you are still choosing an engine, our roundup of AI video models in 2026 covers how the H3 video model stacks up against the field, and the H3 explainer has the pricing and open-weights detail. Either way, the cheapest route to a good clip is still to lock the still first in the freeform generator and spend video credits once.

Tools for this guide

Frequently asked questions

How long can a MiniMax H3 clip be?: MiniMax's model card lists an output duration of 4 to 15 seconds at 24 FPS, with 32 kHz stereo audio. The shorter side is 768 pixels by default, and 2K output is produced by a separate module, H3-Regenerate-2K, which feeds the 768p result plus the original context back through the model rather than running a conventional upscaler. Supported aspect ratios include 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. Two prompt-side consequences follow: every shot timestamp you write must fall inside the duration and increase strictly, and a clip in that range still only holds about one beat per shot, so a second action means a second shot.
Does MiniMax H3 generate audio, and how do you prompt it?: Yes. H3 generates 32 kHz stereo audio jointly with the picture, and the prompt format splits that audio across three places instead of one. Dialogue, singing, and any music or sound a character can actually hear go in integrated_multimodal_description at the moment they happen. Ambient sound, physical action sounds, and non-verbal human sounds such as breathing or laughter go in overall_soundscape, in one to four sentences. Score that only the audience hears goes in non_diegetic_music, in one to three sentences describing instrumentation, tempo, rhythm, and dynamics rather than mood. Write N/A in non_diegetic_music when you want no score; that is H3's equivalent of appending "no music" to a Veo prompt.
How do you write an image-to-video prompt for MiniMax H3?: Image-to-video is the I2VA task, and it starts with a fixed line: "For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced." That line comes first, then one blank line, then the three core fields. Shot 1 should re-establish what the image already fixes, meaning subject, clothing, colors, key objects, and spatial relationships, so the model holds them, and then describe what happens next. MiniMax's recommended shape is first-frame anchor, action onset, continuous development, then result or reaction. Derive the style from the reference image instead of inventing a new one, or the render drifts away from the frame you approved.
Is MiniMax H3 the same thing as the Hailuo app?: Hailuo is MiniMax's consumer video brand and app; H3 is the model behind it, also sold as MiniMax-H3 on the platform API and carried by partners under names including Hailuo 03 and hailuo-3. The prompt-craft difference is what sits in front of the model. Hosted surfaces run H3-Context-IR, a preprocessing system that reads your free-form text and reference files and rewrites them into the structured three-field representation the base model actually consumes. Write that structure yourself and you replace the rewriter's inferences with your own decisions. Specs, pricing, and the open-weights situation are covered in our MiniMax H3 explainer.
Can I run MiniMax H3 locally?: Partly. MiniMax publishes two BF16 checkpoints on Hugging Face, MiniMax-H3 Base FL2VA for text and first/last-frame tasks and MiniMax-H3 Base Ref2VA for reference tasks, under the MiniMax H3 Community License, with deployment paths for SGLang, vLLM, diffusers, and ComfyUI. Two modules are not in that release. H3-Context-IR, the hosted prompt-preprocessing system, is not open-sourced, so nothing rewrites your prompt locally and you write the structured format yourself. H3-Regenerate-2K, which produces the 2K output, is also not yet open-sourced, so a local deployment gives you 768p. MiniMax provides an API for reproducing the full 2K workflow.
How is prompting H3 different from Veo 3.1 or Seedance?: The three dialects are not interchangeable, and pasting one into another fails silently because every model reads unknown syntax as plain text. Veo 3.1 takes a single plain-language paragraph with no token syntax, labels audio inline with SFX: and Ambient noise:, and rewards brevity. Seedance 2.0 takes a six-slot description and tags references numerically, as [Image 1] or @Image1 depending on the surface. H3 takes named fields, timestamped shots, camera motion quantified by amplitude and speed, and stable speaker IDs, and it is the only one of the three that separates diegetic from non-diegetic sound as a formal field. What carries over unchanged: one beat per shot, and dialogue as an exact quoted line rather than a topic.
What do the <d>, <scenetrans>, and <cutoff> tags do?: They are special tokens in H3's tokenizer, not decoration. <d> wraps spoken content: inside it goes only a language tag and the exact words, preserved verbatim including punctuation and never translated, while everything identifying or performative, meaning the speaker ID, voice description, action, and delivery, stays outside the tag. <scenetrans> marks the connecting points on both sides of a cut when one line of dialogue or lyrics carries across it, and it should be paired with an explicit continuity phrase such as "continues seamlessly across the cut". <cutoff> marks speech that the end of the video truncates. For voiceover, use the exact phrase "says in an off-screen voiceover" and state right after the <d> block that the character's lips stay closed.