Demo · Picky — MiniMax H3

Source

Source name

Clip this cut only

Trim in (s)

Trim out (s)

Speed

H3 generates 5–15s, so short cuts are trimmed, never generated. Trim and speed belong to the clip — seconds into this source's lineage, its takes end to end — so they never touch the prompt, and a second clip on the same source can be trimmed differently. You can also drag the clip's edges on the timeline.

Playback sound

Heard as

Level (dB)

Sync (s)

Sound is part of the clip, not a track of its own. A re-roll you rejected for its picture is the same lineage on the same clock, so choosing it here reads its audio over exactly the window the picture plays — trim the clip and the sound follows. Use Sync only if a borrowed take drifts.

then render

DIRECTION

Build the scene

Scene foundation

The one or two sentences that come before [Shot 1]: the style first, the scene it opens on second. The camera line is not part of that opening — it is woven into shot 1, and only when shot 1 has no motion of its own.

Shot LoRAs

LoRAs for this shot

Visual style

Setting & atmosphere

Default camera

Say co-present speakers are different people

Only composes when two speakers share one shot — then it adds “Keep (S1) and (S2) visually distinct throughout…”. Untick it if that sentence reads as noise here.

Storyboard

Each becomes [Shot N] At MM:SS.mmm, …; only the start is emitted. Shots tile the clip end to end, so one shot's end is the next one's start — type into either and both move, and the shot on the other side of the cut gives up exactly what this one gains. Shot 1 starts at 0 and the last shot runs to the budget, so those two ends are shown but not editable. Fill each shot in its written order — composition → subjects, environment, action → camera — and give every shot a camera move of its own (type + amplitude + speed). A cut has to bring new information; if only the distance changes, move the camera instead of cutting.

A speaker key maps to (S1), (S2)… by first appearance across the whole generation; two keys on one line make a group (S1,S2). Only the words and the language go inside <d> — who is speaking, and how, stays outside it.

Frame text & constraints

On-screen text (one per line, written literally)

Negatives not sent

Soundscape — diegetic

Ambient, physical-action and non-verbal human sound across the whole generation, as overall_soundscape. Not dialogue, not the score. Empty means H3 is left to invent the sound, so say something even if it is quiet.

Sound bed

Timed cues

Sound pinned to the same ruler as the shots, appended to the bed as at 6s: …. Leave End blank for an instant hit.

At	End	Sound	Shot

Non-diegetic music

The score only the audience hears. Instrumentation, tempo, rhythm, dynamics — not what it makes you feel. Music a character can hear is a diegetic event and belongs in a shot instead. Empty composes to N/A.

Score

In this generation

Keyframes & references

Keyframes image-to-video

A first frame anchors where the clip starts; an optional last frame is where it resolves. Setting a keyframe renders this segment as image-to-video. Generate writes one from a description instead of uploading a file — the same picture can be reused anywhere, and every image you make is kept in the project's library.

Subjects what the video contains

H3 tracks subjects, not files. A person, a room or a prop is a <Subject N>; the pictures of them are only where it was cut from, and get cited inside its line rather than defined separately. Every cast member on screen who has identity files becomes a subject here on its own — edit the role only to change what this segment says about them. Add one by hand for something with no cast identity behind it.

References reference-to-video

Images, videos or audio the model should honour. Each gets a tag like <Picture 1>. The role becomes its own subject_definitions line — but only when the file stands alone, as a keyframe, a source video or an audio track. A picture a subject is cut from says nothing here: the subject speaks for it, and the row shows which one. Any reference renders this segment as the six-section reference-to-video document. From cast adds a member's identity files as rows so you can set the role and retention for this segment — those rows stay linked, so the file itself still follows the member. A member on screen also sends their files with no row here at all; adding one only takes over the wording.

Task types (ref2v summary prefix)

What the references are for. Composed as the [...] prefix on the summary line, joined with “+”. Tick none and a sensible one is inferred.

Reference summary (ref2v)

Composed prompt

Colour marks where each run of text came from. Click one to jump to the field that wrote it; click a legend chip to mute that kind. Editing this text by hand drops the marks — the words stop being traceable to a field.

Generate (s)

Aspect

Preview

Picky

Project library

Source

Build the scene

Validation

Composed prompt

Usage

Rewrite

Prompt history

Projects

Images

Image prompt history

Take

Monitor

Attach an existing picture