H3 Prompt Composer V5.40.2

✨ AI Project Setup

LLM interchange schema v5

Let an LLM do the first-pass setup, then direct and refine it here.

You can build the whole project at once or work one Generation at a time. For a new project, copy the new-project instructions. After Generation 1 is in the Composer, use Copy next-Generation context, describe only what should happen next, and the LLM can return an append JSON that adds Generation 2 without replacing Generation 1.

Media labels and reusable Picture routing

Project image names are not automatically H3 Picture slots. Use P## only when you truly mean a fixed physical ComfyUI Picture input. Environment views are normally reusable library media: name them by environment + view, then route them into the available Picture inputs for each Generation.

Recommended pattern: keep recurring character/object references pinned when that makes your workflow easier, but let environment views reuse the remaining Picture inputs. Example with four pinned Subjects: Jungle Clearing views can use Pictures 5–6, Jungle Road can reuse Picture 5, and Broken Bridge views can reuse Pictures 5–7. Different Environment nodes may intentionally share those same slot numbers because only the selected Environment is active in a Generation.

P01_Jane_CHARACTER_SHEET.pngFixed/pinned Picture 1 · character identityP02_John_CHARACTER_SHEET.pngFixed/pinned Picture 2 · character identityENV01_JungleClearing_V01_WIDE.pngLogical environment view · no permanent Picture slot impliedENV01_JungleClearing_V02_SHED_SIDE.pngSecond view of the same EnvironmentENV02_JungleRoad_V01_WIDE.pngAnother Environment · may also route to Picture 5ENV03_BrokenBridge_V03_RAMP_DETAIL.pngThird Broken Bridge view · may route to Picture 7 when that Environment is activeG01_JaneJohn_CONT_BLOCKING.pngContinuity image · add P## only if its physical slot is already decidedA01_Jane_VOICE.wavAudio 1 · fixed audio input when knownV01_Performance_REFERENCE.mp4Video 1 · fixed video input when known

When giving files to the AI: label each Environment consistently and distinguish its views (WIDE, APPROACH, SHED_SIDE, RAMP_DETAIL, etc.). The order you upload images or a name like “image 10” is only a project-media identifier unless you explicitly say it is physical Picture 10. After import, verify the Picture chips on each Environment node match what you will actually connect in ComfyUI for that Generation.

Instructions preview

Import AI Setup JSON

Paste the JSON returned by the LLM or open a .json file. Pasted JSON is checked automatically. The JSON’s operation field decides what happens: replace_project replaces the whole project, while append_generations (returned when you use Copy next-Generation context) merges the new Generation(s) into the current project. Structural/schema problems block import; Prompt Check problems are shown clearly but may still be imported so you can fix them inside the Composer.

Camera Language Guide

Build one clear camera instruction. Describe the subject’s action separately, name physical subject sides only when needed, and state camera endpoints instead of relying on screen direction or rotation terminology.

Recommended sentence order

1. Preserve action
Keep recorded performance and path unchanged.

2. Opening camera
State shot size, height, position, and lens aim.

3. Camera move
Name only the camera’s physical path.

4. Ending camera
State the final shot size, position, height, and lens aim.

Useful terminology

Push in / pull back: move the camera closer to or farther from the subject. Truck: move laterally. Pedestal: move the entire camera vertically. Pan / tilt: rotate the lens without moving the camera through space. Arc: move around a subject between named opening and ending positions. Physical right/left: the subject’s own sides, not the viewer’s screen.

Insert a tested phrase

These buttons add concise wording to the active custom camera instruction. Replace bracketed placeholders where shown.

Proven concise pattern

Camera: use a wide, full-body shot of the primary subject with the immediate location visible. Place the camera on the physical right-hand side of the primary subject, measured from the subject’s own perspective, and show that right side in profile. Travel parallel to the primary subject’s recorded path at matching speed. Keep the lens aimed at the primary subject and maintain the selected framing throughout.

Live custom-instruction check

🎞 Video Frame Grabber

local video · native-resolution PNG

Drop source footage or a previous-generation video here

MP4, MOV, or WebM · stays entirely on your computeror

00:00.000

Full-resolution captureThe PNG is drawn at the video's decoded native width × height, not the size of the preview on screen.

Output folder

Browser Downloads (default)

Drop a folder here to save captures directly into it, or choose a folder. If direct folder access is unavailable, Capture PNG uses the browser's normal download behavior.

Picture input #

Source Generation #

Reference purpose

Short name / targets

Keep this compact. Example: JaneJohn for a blocking frame, Jeep for vehicle position, ApartmentWide for a location view. Filename tokens use portable Latin letters/numbers; accents are normalized when possible, while names written only in non-Latin scripts may fall back to “Frame”.

Filename preview

P01_G01_Frame_CONT_BLOCKING_00-00-000.png

Suggested naming keeps the purpose obvious for AI Project Setup. For a rough Photoshop composite, capture the source frame first, edit it externally, then load it as a Spatial Placement Map. It supplies coordinates only, never target pixels or finished integration.

Guided Reference Setup

images · video · audio

Plan the physical Picture, Video, and Audio inputs connected to H3. You may own many more files than the model can receive at once; this wizard defines what each connected input is reserved for. Your work stays open until you explicitly choose Close, Cancel, or Create reference setup.

How many model inputs are you using?Count connected inputs, not every file in your project. H3 accepts up to nine Picture inputs.

I will use Picture inputs

How many Picture inputs are connected?

I will use Video inputs

How many Video inputs are connected?

I will use Audio inputs

How many Audio inputs are connected?

What is each input reserved for?Assign characters, environment banks, scale, continuity, composition, wardrobe, lighting, performance, or audio roles. Shot routing happens later while you build the Generation.

Choose at least one media type above.

Review the input mapThe setup creates reusable library references. Nothing is assigned to a Shot yet.

Keep the new library cards compact after setup

Nothing selected yetChoose the media you want to connect.

Editing Setup Wizard

guided setup

Answer the questions below and the Composer will create the source and reference inputs needed for this edit.

What kind of edit are you making?The available choices are specific to image or video editing.

Edit type

Describe this editThe questions below change with the selected edit type.

Map the model inputsVideo 1 is always the source footage. Choose the physical Picture socket for every additional visual reference.

Replacement subject, object, wardrobe, or appearance referencePicture input

Environment, lighting, or visual-look referencePicture input

Placement or rough-composite guidePicture input

Keep source audio available

Setup readyThe wizard will build the required input cards and open the matching edit controls.

Project setupMode · timing · style

Mode

Timing

Target duration (seconds)

Style

In T2VA / I2VA / FL2VA / L2VA, style opens Shot 1. In Ref2VA, it is written as a brief preamble before Shot 1. e.g. Cinematic, live-action, 2D-animated, 3D CG, claymation, watercolor, vintage film.

Generations — references are shared; appearance continuity is inherited

Shots

Audio

Soundscape status

Use this only when there is no dialogue, ambience, or physical sound in the scene. Non-diegetic music is controlled separately below.

Overall ambience / continuous sound bed — inherited by new Generations

Use this for ambience or recurring physical sound that persists across the Generation. Put one-off impacts, footsteps, engine starts, roars, and other synchronized events inside the relevant Shot.

Additional recurring sounds this Generation — optional

Add only sounds that recur across this Generation. One-off synchronized events belong in the Shot Sound section and do not carry forward automatically.

Music status

This music setting carries forward to a fresh Generation by default. Music heard by characters is diegetic and belongs in the relevant shot, not this field.

non_diegetic_music — 1–3 sentences: instrumentation, tempo, rhythm, dynamics

Prompt check

Ready

Prompt

Check

Suggestions

Phrases

Body: 0 words

Total: 0 / 7,000 chars

Shots: 0

Suggestions run automatically. Evidence-backed H3 adherence checks are combined with lightweight Composer coaching for under-specified actions, repeated Opening/Action information, ambiguous identity wording, timing pressure, and competing reference/camera ownership. Coaching is heuristic and rule-based — it never blocks a prompt or claims to replace an LLM review. Use Locate to jump to the relevant field and dismiss anything that does not apply.

Click to copy. Fragments drawn from MiniMax's own guide examples — use them as register references, not as filler.