MiniMax H3 prompt guide: multi-scene, camera and dialogue

Most H3 prompts are one paragraph of description, and most of them leave the model guessing. The model was built for more than that. It takes a shot-by-shot timeline with camera instructions, speaker IDs, and dialogue blocks, and it ships its own prompt-writing guide in the model repository.

This guide follows that official structure. You will get the three core fields, the shot syntax for multi-scene clips, the camera motion vocabulary, and the dialogue format with voiceover and scene transitions. Everything below comes from the prompt writing skill bundled in the MiniMax H3 repository and the API validation rules.

TL;DR

• An H3 prompt has three fields, in order: integrated_multimodal_description, overall_soundscape, non_diegetic_music H3 prompt writing guide.
• Multi-scene means numbered shots: [Shot 1] with no timestamp, then [Shot 2] At 00:03.500, the camera cuts to... with strictly increasing cut times H3 prompt writing guide.
• Camera motion is type + amplitude + speed, written as a natural action: the camera pushes in with small amplitude at slow speed H3 prompt writing guide.
• Dialogue uses stable speaker IDs like (S1) and text inside <d>[English] ...</d>. Words are preserved verbatim, no translation H3 prompt writing guide.
• The prompt ceiling is 7,000 characters, a character count, not tokens minimax-h3-spec validation rules.
• H3 stably supports dialogue in 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish MiniMax H3 GitHub README.

The three core fields

Every H3 prompt body ends with the same three fields, in this order H3 prompt writing guide:

integrated_multimodal_description: [Shot 1] ...

overall_soundscape: ...

non_diegetic_music: ...

• integrated_multimodal_description is the main body. It carries visuals, actions, shots, speakers, dialogue, singing, and diegetic audio along the timeline. Every detail should correspond to something visible or audible.
• overall_soundscape is 1 to 4 sentences summarizing ambient sound, physical action sounds, and non-verbal human sounds across the whole video. Dialogue and diegetic music already live in the multimodal description, so do not repeat them here. Use N/A only for a requested fully silent video.
• non_diegetic_music is 1 to 3 sentences of background music the characters cannot hear and only the audience can hear. Focus on instrumentation, tempo, rhythm, and dynamics. Not mood words. Use N/A when there is none.

The model repository bundles this as an installable skill, nine skills in total, eight of them style-specific generators MiniMax H3 GitHub README:

npx skills add https://github.com/MiniMax-AI/MiniMax-H3 --skill h3-prompt-writing

It works in any agent that can read local files, no external API calls MiniMax H3 GitHub README.

Multi-scene: the shot structure

A multi-scene prompt is a timeline, not a list. The rules from the official guide H3 prompt writing guide:

1. [Shot 1] opens with no timestamp. State the overall style and initial composition first. Common style anchors: Cinematic, live-action, 2D-animated, 3D CG, claymation, watercolor, vintage film.
2. Later shots use sequential numbers and begin with a strictly increasing cut time inside the video duration: [Shot 2] At 00:03.500, the camera cuts to...
3. Cuts use plain phrasing: the camera cuts to, the shot transitions to, the shot changes to. Cross-dissolve, fade, or wipe only when explicitly requested.
4. Every cut should introduce new information about subject, space, state, viewpoint, or time. If only the distance changes slightly, use camera motion instead of a cut.

A two-shot T2VA example, straight from the guide H3 prompt writing guide:

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 middle-aged baker with a calm, slightly raspy voice (S1) places a fresh loaf on the wooden counter and says: <d>[English] First batch of the morning.</d> [Shot 2] At 00:05.000, the camera cuts to a close-up of steam rising from the sliced bread while the baker's final words carry over from the previous shot.

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.

Notice the cut at 00:05.000 inside a clip that runs to 15 seconds. The cut time is where the new shot begins, not where the previous one ends. That distinction is what keeps multi-scene clips from collapsing into one static take.

Camera motion: type, amplitude, speed

A complete camera expression has three dimensions H3 prompt writing guide:

Amplitude and speed are optional. Medium and normal are usually omitted. Write the motion as a natural English action inside the shot, not as stacked labels at the end of a sentence H3 prompt writing guide:

The camera pushes in with small amplitude at slow speed toward the folded letter in her hands.
The camera pans right with large amplitude at fast speed, revealing the open doorway.
The camera holds a static shot as the runner exits the frame.

Dialogue, speakers, and voiceover

Speakers get stable IDs: (S1), (S2). A speaker keeps the same ID across shots. Characters who never vocalize get no ID. When two numbered speakers talk at once, use a compound ID like (S1,S2) H3 prompt writing guide.

The first time a speaker appears, establish identity from the visual and audio context: character type, age, gender, on-screen or off, pitch, timbre, rate, accent. The spoken words go inside <d> with a language tag, and every original word and punctuation mark stays verbatim. No translation, no rewrite H3 prompt writing guide:

The young woman with a quiet, breathy voice (S1) says: <d>[English] I get off at the next station.</d>

Three rules that catch most dialogue prompts:

• Voiceover uses the exact phrase says in an off-screen voiceover, and the on-screen character's lips stay closed, stated explicitly after the <d> block.
• A line that crosses a cut gets <scenetrans> at the connecting points in both parts, with an explicit note that the audio continues across the cut.
• Speech truncated by the end of the video gets <cutoff>.

On-screen text (banners, signs, subtitles, neon) goes in English double quotes, preserved verbatim: A red neon sign reading "营业中" glows above the doorway. H3 prompt writing guide

The API docs expect prompts in the same three-part shape, subject, action and camera, sound, with (S1) and (S2) speaker tags for dialogue. A prompt that ignores the structure is still a legal request. It just tends to produce worse video minimax-h3-spec validation rules.

Keyframe modes: I2VA, FL2VA, L2VA

When you attach images, the prompt gets an alignment instruction as its first line H3 prompt writing guide:

• I2VA (one image, first frame): For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced. Then anchor style, subjects, composition, and scene in the image before describing what develops forward.
• FL2VA (two images, first and last): 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 N) aligns with the S.SS-second mark of the target video. Describe the path between the frames, not the frames themselves. Single-shot is the default so the model can interpolate continuously.
• L2VA (one image, last frame): the image anchors the final moment. Infer a plausible earlier state, then describe the convergence into the reference image.

The H3-Base-FL2VA checkpoint accepts zero, one, or two input images, which maps to T2V, first/last-frame, and first-and-last-frame workflows MiniMax H3 GitHub README.

Full-reference mode (Ref2VA)

Ref2VA is the multi-reference workflow: up to 9 images, 3 video clips, 3 audio clips, 12 files total MiniMax H3 GitHub README, minimax-h3-spec validation rules. The prompt rewrites into six sections in fixed order: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music H3 prompt writing guide.

Reference labels stay consistent across all sections: <Subject N> for reusable visible content (people, objects, styles), <Picture N> for frame anchors, <Video N> for whole-video structure, <Audio N> for copied audio H3 prompt writing guide. One label, one meaning, from the first section to the last.

One trap is documented in the validation rules: a ref2va request with an empty references block is accepted, billed, and rendered as plain text-to-video. Nothing in the response tells you the references were missing minimax-h3-spec validation rules.

Practical limits

The 7,000-character ceiling is generous. Hitting it usually means something is being concatenated in a loop, not that the prompt ran out of room minimax-h3-spec validation rules.

For hardware context on what these prompts cost to render locally, see the MiniMax H3 VRAM and GPU guide. Duration and resolution constraints that shape how you pace a multi-scene clip are in how many seconds H3 really generates. The ComfyUI setup for running these prompts locally is in the MiniMax H3 ComfyUI guide.

Common errors and fixes

FAQ

How to prompt MiniMax H3? Use the three-field structure: integrated_multimodal_description with numbered shots and cut times, overall_soundscape in 1-4 sentences, non_diegetic_music in 1-3 sentences. Keep it under 7,000 characters H3 prompt writing guide, minimax-h3-spec validation rules.

Can you provide some examples of MiniMax H3 prompts? The model repository ships a prompt-writing skill with worked cases for T2VA, I2VA, FL2VA, and L2VA, including a two-shot bakery scene with dialogue and a cut at 00:05.000 MiniMax H3 GitHub README, H3 prompt writing guide.

What are some tips for MiniMax H3? State style and composition in Shot 1, give every cut a strictly increasing timestamp, write camera motion as type + amplitude + speed inside a natural sentence, and keep dialogue verbatim inside <d> tags with stable speaker IDs H3 prompt writing guide.

How does MiniMax H3 work? It is an omni-modal system that takes unified text, image, video, and audio context and generates video with native stereo audio up to 15 seconds at 2K. The prompt structure exists because the model is trained to follow shot-level timelines MiniMax H3 launch post, MiniMax H3 model card.

Can MiniMax H3 do multi-scene videos in one generation? Yes, up to 15 seconds. Use numbered shots with cut times inside the duration. For longer stories, generate 8-second segments and stitch, since 8 seconds is the only grid-exact duration H3 prompt writing guide, minimax-h3-spec validation rules.

Which languages does H3 dialogue support? Stable support for 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. Additional languages work to varying degrees MiniMax H3 GitHub README.

Sources

• MiniMax-AI/MiniMax-H3 (GitHub) — Model specs (24 fps, 32 kHz stereo, 11 dialogue languages), FL2VA/Ref2VA input limits, installable h3-prompt-writing skill.
• H3 prompt writing guide (base-en.txt) — Three core fields, shot/cut syntax, camera motion table, speaker IDs, dialogue and voiceover rules, worked T2VA/I2VA/FL2VA/L2VA cases.
• minimax-h3-spec: Validation rules — 7,000-character prompt limit, expected prompt shape, Ref2VA silent degradation.
• MiniMax API: Video Generation — Duration and resolution parameters, reference input limits.
• MiniMaxAI/MiniMax-H3 (Hugging Face) — Model card: omni-modal context, 15s at 2K, native stereo audio.
• MiniMax H3 launch post — Official capability summary: complex multimodal instructions, stereo audio, 2K.