MiniMax H3 Prompt Guide: 5 Modes and the Six-Section Structure

Updated 2026-09-20 · Sources: ComfyUI's official H3 prompt guide, MiniMax's base-mode writing guide and full-reference guide, minimaxh3.co and minimaxh3.cc

Why format beats adjectives

H3 generates picture and sound jointly, so a single prompt has to place camera moves, dialogue, sound effects and music into their own slots. As ComfyUI's official guide puts it, H3's multimodal training pays off exactly here: shots, camera moves, dialogue and sound effects all live in one prompt block. The prompt is therefore a timeline — what happens at which second, who speaks which exact line, what keeps sounding underneath, and when the score fades. Put one of those layers in the wrong field (say, dialogue inside the ambience field) and the model receives two competing sets of instructions.

MiniMax publishes two writing guides: one for the base modes (T2VA, I2VA, FL2VA, L2VA) and one for the full-reference mode (R2V, often written Ref2VA). The base tasks share one three-field audiovisual body, with a precise alignment line prepended for image-anchored tasks. Full-reference is not a longer base prompt: it first defines what each reference asset contributes, then records how that contribution is used along the target timeline. Classify the task before writing a single visual detail and most rewrites disappear (sources: the ComfyUI guide page and the two official MiniMax guides).

The five modes compared

Table sources: MiniMax's official base guide (inputs and alignment rules for T2VA/I2VA/FL2VA/L2VA) and full-reference guide (the R2V six sections); community field guides at minimaxh3.co / minimaxh3.cc. In the alignment lines, N is the index of the actual final shot and S.SS is the clip duration with exactly two decimals; the instruction must be the first line, followed by one blank line.

The three core fields, explained

All four base tasks share the same audiovisual body: three fields, each owning one information layer. Writing them as one flowing paragraph is the most common source of failures in H3 prompting.

integrated_multimodal_description: [Shot 1] ...

overall_soundscape: ...

non_diegetic_music: ...

• integrated_multimodal_description — the visual timeline, the main body. Describe everything visible and audible over time. Open with the overall style and initial composition (Cinematic / live-action / 2D-animated / 3D CG / claymation / watercolor / vintage film, and so on), then cover subject appearance and position, scene and key props, actions and reactions, cuts, speakers and dialogue, and synchronized in-scene sound. Shot rules: [Shot 1] carries no timestamp; later shots begin like [Shot 2] At 00:03.500, the camera cuts to …, with strictly increasing cut times inside the requested duration. Add a cut only when the next shot introduces new subject, space, state, viewpoint or time information; for a distance or angle change, use camera motion instead (Push In / Pan / Truck / Arc / Tracking / Static, with with small amplitude or at slow speed when they matter), and write it as a natural action sentence rather than stacked labels.
• overall_soundscape — full-clip ambience, 1–4 sentences. Summarize ambient sound, physical action sounds and non-verbal human sounds: wind, rain, street tone, footsteps, fabric movement, impacts, breathing, laughter. Do not repeat dialogue, singing or on-screen music here — those belong in the timeline — and never put the score here. Use N/A only when complete silence is explicitly required.
• non_diegetic_music — audience-only score, 1–3 sentences. Describe music the characters cannot hear: instrumentation, tempo, rhythm and dynamic change. Do not replace instrumentation and tempo with abstract mood words. Write N/A when no score is wanted. Radio, television, on-screen instruments and singing audible to the characters are diegetic events and stay in the timeline instead.

Speakers and dialogue: give every actually vocalizing source a stable ID (S1, S2; two speakers together can be (S1,S2)) and reuse it across shots. Identity, voice, action and delivery stay outside <d>; inside it, keep only the language tag and the exact words, e.g. The young woman with a quiet, breathy voice (S1) says: <d>[English] I get off at the next station.</d>. Voiceover must be stated as an off-screen voiceover with the on-screen character's lips staying closed; use <scenetrans> when a line continues across a cut and <cutoff> when the video ends mid-line. Characters who never vocalize get no ID.

On-screen text: any text that must be visible (signs, banners, subtitles, neon) goes inside English double quotation marks, verbatim and untranslated. Community testing cited in ComfyUI's guide notes that enumerating every string in quotes keeps it legible, while undefined text tends to collapse into jumbled symbols.

ComfyUI extras: H3 supports the standard embedding: syntax — drop a style embedding into ComfyUI/models/embeddings/ and reference it by name, e.g. embedding:my_embedding. The Comfy-Org/MiniMax-H3 repository hosts 10 style embeddings (contributed by community member silveroxides; unofficial). (Source: ComfyUI's official prompt guide page.)

Ref2VA and the six-section structure

Full-reference mode (Ref2VA / R2V) is not a longer three-field prompt: it first defines what each reference asset is and contributes, then records how it is used and retained along the target timeline. All six sections must be present in order; unused labels can be removed, required sections cannot.

1. subject_definitions — define every reusable subject, concrete frame anchor, whole-video source and audio source. For example: "<Subject 1> is the woman in <Picture 1>, preserving her red coat and facial identity."
2. summary — one short paragraph with the task type (keyframe completion / reference generation / video editing / video continuation / audio reuse / audio reference, combinable) and the main reference relationships.
3. retention_analysis — record how each visual or audio reference is preserved, transferred, copied or loosely followed, naming the concrete features.
4. detailed_description — write the target video shot by shot in playback order; the overall style goes before [Shot 1] (the opposite order from base mode), and each label is cited where its role actually takes effect.
5. overall_soundscape — ambience and physical sounds, including any applicable audio copy/reference relationship.
6. non_diegetic_music — the audience-only score and its relationship to referenced audio, or N/A.

Retention markers: on the visual side, fully_preserved (keep the full defining role — name the features that stay unchanged), partially_preserved (keep it but alter or omit part — say exactly what changes), attribute_transfer (move a recognizable property such as motion, style, pose, composition or effect onto a different subject), and weak_reference (follow only a broad category, atmosphere or stylistic resemblance). On the audio side, fully_copy (the entire source signal becomes the final track), partially_copy (copy selected time ranges or layers, or add/remove/replace other sounds), and reference (follow timbre, delivery, beat, music style, dialogue content or texture without copying the signal).

Three pitfalls worth stating outright. First, get the audio marker right: ComfyUI's guide notes that fully_copy uses the source audio 1:1 as the final track, while reference only carries timbre and delivery — with the wrong marker, H3 treats the audio as a timbre reference and lip movements drift out of sync with the lyrics (community-verified). Second, the official guides require all six section bodies in English, keeping only exact dialogue, lyrics and visible on-screen text in their original language. Third, for a typical generation task, 350–500 English words is a useful detail range for detailed_description — not a quota; dialogue-heavy clips should prioritize the complete spoken timeline. Also note that reference audio cannot be the sole input: it must accompany image or video material (source: minimaxh3.cc prompt library).

Two complete example prompts (copy and adapt)

Example 1 · T2VA (with dialogue and ambience, ~5 s):

integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium-wide shot frames a bakery owner rolling up the shutter of a small corner bakery before sunrise. The camera pushes in with small amplitude at slow speed as he sets the first loaf on the wooden counter and says in a calm, low voice (S1): <d>[English] First one of the morning. Take it while it's warm.</d> [Shot 2] At 00:04.000, the camera cuts to a close-up of steam rising from the sliced bread while his final words carry over from the previous shot.

overall_soundscape: The metal shutter rattles to a stop, leaving a quiet street tone and the faint simmer of a pot behind the counter. The loaf meets the wood with a soft thud, followed by the light click of a bread knife.

non_diegetic_music: Sparse piano notes at a slow tempo, joined by a sustained low string tone that gradually fades out.

Example 2 · I2VA (with the first-frame alignment line):

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, cinematic, the cyclist in <Picture 1> waits under the convenience-store awning, preserving her rain jacket, her standing position and the street composition of the image. The camera trucks right with small amplitude at slow speed as she looks up at the pooled water on the road, tilts her folded umbrella forward, and says in a slightly hoarse young voice (S1): <d>[English] One more minute and it lets up.</d>

overall_soundscape: Rain taps steadily on the awning fabric; tires hiss through water on the road below, and a neon sign hums faintly somewhere behind her.

non_diegetic_music: N/A

Note: examples are written in English because the official full-reference guide requires all six section bodies in English (only dialogue, lyrics and visible on-screen text keep their original language); field names and alignment lines must stay in the original English wording in every mode.

Before generating, run the official checklist: is the alignment instruction the first line, followed by one blank line; does S.SS use exactly two decimals and [Shot N] name the actual final shot; are cut times strictly increasing and inside the duration; do speaker IDs stay stable across shots with only the language tag and exact words inside <d>; does overall_soundscape avoid repeating dialogue or score; does non_diegetic_music describe instrumentation, tempo, rhythm and dynamics — or correctly say N/A?

Five common mistakes

1. Putting dialogue into overall_soundscape. The soundscape field holds ambience, physical sounds and non-verbal human sounds only; dialogue, singing and on-screen music belong to the timeline. Repeating one line in two fields gives the model competing audio instructions.
2. Mixing score with in-scene sound. Score the characters cannot hear goes only into non_diegetic_music, described by instrumentation, tempo, rhythm and dynamic change. A radio, on-screen instrument or singing the characters can hear is a diegetic event and stays in the timeline. If you want silence, write N/A — don't just omit the field.
3. Camera directions that fight the action. Cutting a [Shot 2] for a small push-in, or moving the camera against the subject's motion, fragments the clip or breaks continuity. Cut only for new information; use camera motion for distance and angle changes; keep one clear camera intent per shot.
4. Forgetting N/A. Leaving non_diegetic_music blank invites the model to add a score of its own — write N/A when no audience-only music is wanted. In overall_soundscape, N/A is reserved for explicitly requested complete silence; don't swap the two.
5. Reference roles left vague. Writing "keep consistent" without naming identity-defining features, choosing a retention marker and citing labels in the timeline produces output that is broadly similar but drifts where it matters. Instead, name the features that define success in subject_definitions, pick the right marker (fully_preserved / partially_preserved / attribute_transfer / weak_reference), and cite each label in every shot where it takes effect.

Prompts don't change hardware needs

No matter how long the prompt or how many references you attach, the compute bill is still set by resolution × duration (plus model, precision, step count and attention backend). Two hard edges from ComfyUI's documentation are worth remembering: H3's native canvas is a 768px short edge — 1344×768 at 16:9, with resolutions rounded to a multiple of 32 — and the duration snaps to the 17-frame-per-block (17k+5) grid at 24fps. In other words, a 2,000-character prompt and a 200-character prompt cost exactly the same time and VRAM at the same resolution and duration.

Estimate your own runtime with this site's hardware calculator (a resolution × duration matrix estimate; see the about page for methodology). To bring the numbers down, start with which LoRA acceleration actually fits you — solve compute first, polish prompts second.

Sources

• ComfyUI official MiniMax H3 prompt guide (general tips, on-screen-text quoting rule, embedding support and the 10 community style embeddings): docs.comfy.org
• MiniMax official base-mode writing guide (T2VA / I2VA / FL2VA / L2VA structures and cases): VIDEO_PROMPT_WRITING_GUIDE_base_en.md
• MiniMax official full-reference writing guide (R2V six sections, labels, retention markers): VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
• Official H3 skills, including the prompt-writing skill: github.com/MiniMax-AI/MiniMax-H3
• Field guide and failure diagnosis (three-field and six-section templates, common failures with fixes): minimaxh3.co/prompt-guide
• Prompt library (10 prompt skeletons and worked examples; the Ref2VA example follows the official structure): minimaxh3.cc/prompts

No numbers or claims on this page are left unattributed; if the official guides change, the Hugging Face and ComfyUI originals win.