Skip to main content
H3 expects shots, camera, dialogue, and SFX in one multimodal prompt. This page collects official guides, cross-workflow tips, shot/dialogue rules, and style embeddings.

​
Official guides

  • Base modes (T2VA, I2VA, FL2VA, L2VA): timed shots, camera, audio
  • Full-reference mode (R2V): rewrite structure, reference labels, retention markers
Installable H3 skills package these for agents. Applied per template: T2V/I2V, R2V, Multiframe, Fun ControlNet.

​
General tips

I2VA opens with a fixed image-alignment line, blank line, then body fields. FL2VA/L2VA use their own alignment line; T2VA starts at the fields directly:
Write prompts in English. Dialogue/lyrics in <d> tags; on-screen text verbatim in its original language.
  1. Scene first: Overall scene, then timed shots inside [Shot 1] of integrated_multimodal_description
  2. One block: Shots, camera, and audio together
  3. Resolution: Native 1344×768 at 16:9; multiples of 32 (details)
  4. Length: length is frames at 24fps (default 124 ≈ 5s), snapped to 17k+5
  5. On-screen text: English double quotes per string, verbatim (e.g. A red neon sign reading "营业中"). Spoken lines use <d>, not quotes
  6. R2V audio markers: In retention analysis, mark each audio ref as fully_copy, partially_copy, reference, or weak_reference to match how the clip is used (see official ref guide table)
  7. Audio fields: overall_soundscape = ambient/SFX/non-verbal (1–4 sentences; N/A only if truly silent). Dialogue and diegetic music stay in shots. non_diegetic_music = score only (1–3 sentences; N/A if none). If unwanted speech appears, fill both fields explicitly
  8. No negatives: BasicGuider runs at cfg 1 with no negative branch. Ban lists add unwanted concepts. State what should appear (or use R2V retention fields: fully_preserved, partially_preserved, etc.)
  9. Reference sheets (R2V): One image can hold multiple panels. Give the sheet its own <Picture N> (not only inside <Subject N>), describe panels, and point shots at a panel (e.g. full front view of <Picture 2>). Label panels clearly
Workflow-specific tags and controls live on each workflow page.

​
Shots and cuts

Number shots [Shot 1], [Shot 2], … [Shot 1] starts at 0.00s. Later shots open with a cut time:
  • Cut times increase and stay within clip duration (length at 24fps)
  • Use a single timestamp (At 00:05.000), not a range like [5s-8s]
  • Phrase cuts as actions: the camera cuts to, the shot transitions to, etc. Dissolves/fades/wipes when needed
New shot when subject, space, state, viewpoint, or time changes. Small reframes stay one shot with camera motion.

​
Dialogue and speakers

  • Speaker IDs: (S1), (S2), … for every voice (including off-screen/singing). IDs index speakers, not character numbers. Duet: (S1,S2)
  • <d> content: ID, action, and delivery outside tags; inside tags, language tag + spoken words verbatim, in the shot paragraph where spoken
  • Punctuation in <d>: Keep , . ? !; drop decorative repeats/emoji; end statements with . ? ! before </d>; use [unclear] when unsure. Ellipses often hurt intelligibility
  • Voiceover: Use says in an off-screen voiceover and state lips stay closed after </d>. Cross-cut dialogue: <scenetrans> at both sides plus audio carries over
  • Multi-speaker + audio refs: Wrong-speaker assignment can happen. Tie lines to visible events, not timecodes. If needed, generate per-speaker audio externally and reference with the right marker (fully_copy vs partially_copy)

​
Prompt embeddings

ComfyUI#15697 supports embedding:name in H3 prompts. Place files in ComfyUI/models/embeddings/. Comfy-Org/MiniMax-H3/embeddings lists 10 community style embeddings (silveroxides, discussion #50; unofficial). Trigger word = filename without extension: