ComfyUI MiniMax H3 Prompt Guide: Official Guides, Tips, and Prompt Embeddings

​

Official guides

• Base generation modes (T2VA, I2VA, FL2VA, L2VA): structure a prompt into timed shots with camera movement and audio (dialogue, SFX, music), with examples for each mode
• Full-reference mode (R2V): the rewrite output structure, including subject definitions, reference labels, and retention analysis, and how to assign each reference a role in the target shot

​

General tips

integrated_multimodal_description: [Shot 1] ...
overall_soundscape: ...
non_diegetic_music: ...

1. Describe the whole scene: State the overall scene first (location, character, what is happening), then break it into timed shots. The overall style and the opening composition belong at the start of [Shot 1] inside integrated_multimodal_description, after any image-alignment instruction.
2. Shots, camera, and audio: Describe the shots, camera moves, and the accompanying audio (dialogue, SFX, music) in one prompt block
3. Resolution: H3’s native canvas is a 768px short edge, which is 1344x768 at 16:9, and resolutions are rounded to a multiple of 32. See Setting the output resolution
4. Length: The node’s length input is a frame count at 24fps (default 124, about 5 seconds) that snaps up to the model’s 17-frame block grid (17k+5)
5. On-screen text in quotes: Wrap any text that is visible on screen (signs, banners, subtitles, neon text) in English double quotation marks, and list each string separately. Copy the original text and punctuation verbatim, without translating it, for example A red neon sign reading "营业中" glows above the doorway. Keep spoken dialogue in <d> tags instead: quotes are for printed text, <d> is for what characters say
6. Audio reuse markers in R2V: In full-reference (R2V) prompts, declare each audio reference’s relationship in the retention analysis: fully_copy uses the source audio 1:1 as the final audio track, partially_copy copies part of the timeline or selected layers, or adds, removes, and replaces sounds afterwards, reference carries timbre, rhythm, music style, dialogue content, or sound texture without copying the signal, and weak_reference keeps only a broad similarity in category or atmosphere. The marker must match what the audio actually does in the target video. See the official full-reference guide’s relationship marker table
7. Write the audio lanes explicitly: Each prompt carries two audio fields, and they follow different rules. overall_soundscape summarizes ambient sound, physical action sounds, and non-verbal human sounds in 1 to 4 sentences (dialogue, singing, and diegetic music belong in integrated_multimodal_description, not here), and uses N/A only when the clip should be completely silent. non_diegetic_music describes score that only the audience hears in 1 to 3 sentences, and uses N/A when there is no such music. If a quiet shot comes back with speech you did not ask for, write out both fields explicitly and regenerate. See the official base guide’s sections on overall_soundscape and non_diegetic_music
8. Write bans as positives: the H3 templates sample through the BasicGuider node, which carries a single conditioning input, so there is no negative branch and a negative prompt has no effect (the guider runs at cfg 1, where ComfyUI skips the unconditional pass; a CFGGuider only uses negative conditioning above cfg 1). An instruction that names an unwanted element adds that wording to the description the model reads, with no separate pass to subtract it, so a line such as no subtitles and no on-screen text registers text as content. Say what should be in the shot instead, for example the sign above the door is blank. In full-reference prompts, scope the same thing through the retention analysis, where fully_preserved, partially_preserved, attribute_transfer, and weak_reference state what each reference contributes to the target video
9. Reference sheets (R2V): one reference image can hold several views of a subject or the panels of a storyboard, and the prompt then addresses one panel at a time. Give the sheet its own <Picture N> entry rather than citing it inside a <Subject N> definition, since a picture cited only inside a definition is not used as a separate reference, and describe the sheet there, for example <Picture 2> is a character sheet with three panels: a portrait, a full front view, and a full back view of the woman. Point a shot at one panel inside the shot description, for example [Shot 2] the shot's keyframe corresponds to the full front view panel of <Picture 2>. A sheet keeps several views inside the local R2V node’s limit of 9 reference images and lets each shot name the view it needs. Number or label the panels so the prompt can point at them, and keep the sheet’s printed text to those labels.

​

Dialogue and speakers

• Stable speaker IDs: give every voice in the prompt an ID such as (S1) or (S2), including off-screen and singing voices, and repeat that ID in every shot where it speaks. The IDs number speakers, not subjects: a character who never vocalizes gets no ID, and the first voice in the prompt is (S1) even when its character is numbered differently. Speakers who vocalize together take a compound ID such as (S1,S2)
• Outside and inside <d>: the identifying phrase, the speaker ID, the action, and the delivery stay outside the tags, and the tags carry only the language tag plus the words themselves, copied verbatim. Write each line inside the shot where it is spoken, in the same paragraph as that shot’s action
• Voiceover: a line the audience hears without the character speaking it needs the exact phrase says in an off-screen voiceover, plus a statement right after the <d> block that the on-screen character’s lips remain closed. A line that continues across a cut needs <scenetrans> at both connection points and a statement that the audio carries over the transition
• Several speakers with audio references: in full-reference (R2V) prompts that pair two or more speakers with audio references, H3 can hand an audio reference to the wrong speaker even when the prompt text and the reference connection order are both correct. Keep each shot to the references that shot needs, and tie a line to a visible event, for example When the phone is at his ear, the man in the coat speaks, rather than to an absolute timecode, which is what cut times are for. When a line still lands on the wrong speaker, generate it with a voice tool and supply it as that speaker’s audio reference, with the relationship marker that matches the clip: fully_copy when the clip becomes the video’s complete final audio track, or partially_copy when it covers part of the timeline

​

Prompt embeddings