🎬

MiniMax-H3 Prompt Studio

Build H3-Context-IR prompts that H3-Base was actually trained on.

Task

Duration

8.0 s

Shots

Shot 1 carries the style and opening composition and takes no timestamp. Every later shot needs a strictly increasing cut time inside the duration.

overall_soundscape

1–4 sentences. Ambience, physical action sounds, non-verbal human sounds. Dialogue, singing and diegetic music belong in the description, not here. N/A only for deliberate total silence.

non_diegetic_music

1–3 sentences. Instrumentation, tempo, rhythm, dynamics. No abstract mood words and no explaining the emotional function. Anything the characters can hear is diegetic and belongs in the description. N/A if there is none.

Presets

The worked examples from the official guides.

Prompt


        

      


      

        
Checks

        
    
      
    

      
    
        
    Where this goes
    
        
    The studio builds the prompt; it does not generate video. Feed the output to any
           H3 deployment.
    
        
      
          
    • multimodalart/minimax-h3
              β€” the open weights, unquantized, on ZeroGPU
      • 
          
    • A local SGLang / vLLM / diffusers deployment of H3-Base
              (4 GPUs, BF16)
      • 
          
    • The MiniMax platform API video-generation-v2-create β€” the only route to
              H3-Regenerate-2K 2K output, which is not open-sourced
      • 
        
    
      
    
    
    

  
    

    



    
  
    
    
    Official reference outputs
    
    
    Straight from the model repo. These are what a full H3-Context-IR prompt
       produces β€” the structured form is what the checkpoint was trained to consume.
    
  
    
  

    



    
  
    

    
    
      
    Camera motion
    
      
    Motion type + amplitude + speed, written as a natural English action inside the shot β€”
         not stacked as labels at the end of the sentence. Medium amplitude and normal speed are omitted.
    
      
    
    
    

    
    
      
    Frame snapping
    
      
    H3's video VAE decodes chunks of 17 frames plus a 5-frame head, at 24 fps. Any requested duration is
         snapped up to the next frame count where frames ≑ 5 (mod 17).
    
      
    15.0 s is 360 frames, which snaps to 362 β€” 15.083 s, past the ceiling,
         and the released checkpoint refuses it. The largest valid count under 15 s is 345
         frames (14.375 s).
    
      
    
    
    

    
    
      
    Dialogue and speakers
    
      
      
        
    • Stable IDs β€” (S1), (S2); joint delivery is (S1,S2). An ID
            persists across shots. Characters who never vocalize get no ID.
      • 
        
    • On first appearance, establish identity: type, age, gender, on- or off-screen, pitch, timbre, rate,
            accent.
      • 
        
    • Identity, ID, action and delivery go outside <d>. Only the language tag
            and the verbatim spoken content go inside.
      • 
        
    • Preserve every original word and punctuation mark. Do not translate.
      • 
        
    • Voiceover uses the exact phrase says in an off-screen voiceover, and every voiceover
            <d> must be followed by a statement that the character's lips remain closed.
      • 
        
    • <scenetrans> marks a line crossing a cut, at both connecting points.
            <cutoff> marks speech truncated by the end of the video.
      • 
      
    
    
    

    
    
      
    The three core fields
    
      
      
        
    • integrated_multimodal_description β€” the body. Visuals, actions, shots, speakers,
            dialogue, singing, and diegetic audio along the timeline. Every detail must correspond to something
            visible or audible.
      • 
        
    • overall_soundscape β€” 1–4 sentences, one paragraph. Ambience, physical action sounds,
            non-verbal human sounds. Never repeats dialogue.
      • 
        
    • non_diegetic_music β€” 1–3 sentences. Score only the audience hears. Instrumentation,
            tempo, rhythm, dynamics.
      • 
      
    
    
    

    
    
      
    On-screen text
    
      
    Any banner, sign, label, subtitle or neon text actually visible on screen goes in English double
         quotation marks, verbatim and untranslated.
    
      
    A red neon sign reading "θ₯业中" glows above the doorway.
    
    
    

    
    
      
    Generation constraints
    
      
    
        
    Duration4–15 s
    
        
    Frame rate24 fps
    
        
    Frame count≑ 5 (mod 17)
    
        
    Short edge768 px default; 2K via H3-Regenerate-2K
    
        
    Audio32 kHz stereo
    
        
    Aspect21:9, 16:9, 4:3, 1:1, 3:4, 9:16 and more
    
        
    Dialogue languages11 stable: ar, zh, en, fr, de, it, ja, ko, pt, ru, es
    
        
    GuidanceCFG-distilled β€” no negative prompt
    
      
    
    
    

  
    

    



    
  
    
    
    The module that isn't open
    
    
    MiniMax-H3 ships as three modules. Only the middle one is open-sourced.
    
    
    
      
    ModuleRoleReleased
    
      
    H3-Context-IRTurns free-form multimodal input into the structured representation H3-Base consumesNo β€” hosted API
    
      
    H3-Base33B dense omni-transformer. Generates 768p video + stereo audioYes
    
      
    H3-Regenerate-2KRegenerates the 768p result in-context at 2KNo β€” hosted API
    
    
    
    
    The model card does not treat the first one as optional:
    
    
    H3-Context-IR is critical to the quality of the final output, so we strongly recommend
      incorporating it into your generation pipeline or following the "Prompting Guidance" to build your own
      context-processing system.
    
    
    That is the gap this Space fills. Run the open weights with a one-line prompt and you have skipped the
       module that was in the loop during training β€” H3-Base expects the structured form, and the reference
       outputs in the repo were produced with it. The two official prompting guides specify that form precisely
       enough to build against, which is what the studio does: it assembles the mechanical parts of the format
       and enforces the rules the guides state.
    
    
    What it will not do
    
    
    This is a formatter and a linter, not a rewriter. Real H3-Context-IR does instruction
       parsing, cross-modal association, temporal understanding and logical reasoning across a multi-stage
       hosted workflow, and it will supplement underspecified detail on its own. The studio does none of that β€”
       you write the content, it gets the shape right. For the official behaviour, the
       video-generation-v2-h3-context-ir endpoint reproduces it.
    
    
    Privacy
    
    
    Entirely client-side. No backend, no key, no telemetry; nothing you type leaves the page. The reference
       videos on the gallery tab stream directly from the model repo on huggingface.co.