h3-ref-prompt

# MiniMax H3 — full-reference prompt (with reference media) Goal: turn the user's idea and attached media into a **six-section full-reference brief** for MiniMax H3. Output is only the brief text in English (original language survives only inside `<d>` tags and in on-screen text). If there is NO media — use the `h3-t2v-prompt` skill instead. ## Step 0. Inputs - List the media and their **roles** (ask the user if unclear): first-frame anchor / last-frame anchor / keyframe / general reference (character, scene, style, voice timbre, soundtrack, source video). - **Media order is semantic**: labels are numbered by input order, independently per category — `<Picture 1..N>`, `<Video 1..N>`, `<Audio 1..N>`. Never rename, skip, renumber, or reorder. - Ask in one message what's missing: duration (4–15s, this user defaults to 15), orientation (16:9 / 9:16), speech and its language. - Audio can never be the only media — only alongside an image or video. ## Output format (exactly six sections, in this order) ```text subject_definitions: ... summary: ... retention_analysis: ... detailed_description: ... overall_soundscape: ... non_diegetic_music: ... ``` ### 1. subject_definitions One line per label: what it denotes, its role, the key features to follow. - `<Subject N>` — reusable visible content: people, animals, objects, scenes, wardrobe, props, styles, actions, poses. A content unit, not the source file. One subject may combine several assets: `<Subject 1> is the woman whose appearance comes from <Picture 1> and whose walking motion comes from <Video 1>.` - `<Picture N>` — a standalone line ONLY if the image itself serves as a concrete frame (first/last frame, keyframe, composition anchor) or storyboard. If an image only defines a character/style, cite it inside the `<Subject N>` definition instead of creating a separate line. - `<Video N>` — only for whole-video relationships: the edited source, a continuation start point, a reference for structure/editing/rhythm. Visible content reused from a video still goes under `<Subject N>`. - `<Audio N>` — copied or referenced audio: voice timbre, music style, a copied track. If bound to a speaking subject: `<Audio 1> is the voice-timbre reference for <Subject 1> (S1).` ### 2. summary One short paragraph. Opens with a task-type prefix combined with ` + ` (no repeats): - `keyframe completion` — an image serves as a concrete frame (first/last/keyframe). - `reference generation` — an asset guides character/scene/style/motion/storyboard without being a concrete frame. - `video editing` — a source video is directly modified. Then continue with: `The target video is an edited version of <Video 1>.` - `video continuation` — new content continues a source video. - `audio reuse` — the audio signal is copied in full or in part. - `audio reference` — the signal is not copied; only style/timbre/rhythm is referenced. Include a type only if an asset genuinely plays that role. A video used only for motion/rhythm = `reference generation`, not editing/continuation. Introduce no new labels here. ### 3. retention_analysis One line per label. Fixed markers: - Visual: `fully_preserved | partially_preserved | attribute_transfer | weak_reference` Format: `<Subject 1> (appears in [Shot 1], [Shot 3]): fully_preserved - identity, hair, and pink shirt retained.` - Audio: `fully_copy | partially_copy | reference | weak_reference` Format: `<Audio 1>: reference - the target speaker follows <Audio 1>'s timbre without copying the signal.` - No `(Sx)` in this section. New actions/backgrounds in the target video are not a loss of reference fidelity. ### 4. detailed_description — the body (350–500 words for generation) - Style — 1–2 sentences BEFORE `[Shot 1]`, including orientation: `The target video is in a cinematic live-action style with warm lighting, vertical 9:16 framing.` - `[Shot 1]` has no timestamp; later `[Shot N] At MM:SS.mmm, the shot cuts to ...` — strictly increasing times within the duration. - At the first appearance of a `<Subject N>`, describe its referenced features, position in frame, and current action; afterwards reuse the label without redefining. Name wardrobe in text even for referenced subjects — H3 drifts wardrobe. - Frame anchors in natural phrasing: `the shot begins from <Picture 1>`, `the shot's keyframe corresponds to <Picture 2>`, `the shot ends on <Picture 3>`. - Anchor roles: - **First-frame**: first-frame anchor → action onset → continuous development → result/reaction. Describe the motion LEAVING the frame; do not re-describe the image. - **First+last**: describe the interpolation path between the frames (a SINGLE shot is preferred); the final shot must land exactly on the last frame: first-frame state → observable intermediate changes → progressively narrowing differences → last-frame state. - **Last-frame only**: infer a plausible preceding state and converge onto the image in the final shot. ### Dynamism / smoothness / aliveness rules (user priority) - The timeline is consecutive beats, each with ONE primary change and an observable end state (something a viewer could point at). - The most important beat goes in the MIDDLE of the timeline (the model squeezes the ending). - A complex action (hand-off, pose change) takes ~4 seconds. Doesn't fit — cut the least important one. - A cut = new information (subject/space/state/viewpoint/time). Just getting closer — use camera motion. - Smoothness: continuous action development inside the shot; continuity across cuts — carrying motion, sound `continues seamlessly across the cut`, character reactions. - Emotion → observable: where the eyes go, what the hands do, breathing, what stays still. Micro-detail: fabric, hair, steam, glints, background life. ### Camera — always explicit One move per shot, as a natural sentence: `The camera pushes in with small amplitude at slow speed toward her hands.` Vocabulary: `Zoom In/Out`, `Push In/Pull Out`, `Pan Left/Right`, `Truck Left/Right`, `Tilt Up/Down`, `Pedestal Up/Down`, `Arc Shot`, `Tracking Shot`, `Static Shot`, `Shake Slightly/Strongly`, `POV`, `Roll Clockwise/Counterclockwise`. Amplitude `with small/large amplitude`, speed `at slow/fast speed` (omit medium/normal). Static — `the frame never moves` plus the list of forbidden moves. ### Speakers and audio - A speaking referenced subject: `<Subject 2> (S1) turns and says, <d>[English] ...</d>`. `(Sx)` is assigned in the order of actual vocal events and reused; off-screen speech by the same subject keeps the form plus `off-screen`. - A voice inside a directly reused soundtrack — the source is `<Audio N>`; do NOT invent an `(Sx)`. - With `audio reference` (timbre/rhythm only) — do NOT carry the reference audio's original words into the target. With direct reuse of lines — verbatim, unintelligible spans `[unclear]`, punctuation `, . ? !`. - Voiceover: `says in an off-screen voiceover` + `while his lips remain completely closed.` A line across a cut — `<scenetrans>` at both ends + a continuity phrase. Truncated by the video end — `<cutoff>`. End of speech — lips close, articulation ceases. - `<d>` languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish. - On-screen text — in double quotation marks, verbatim; if no text should appear, say so explicitly. ### 5. overall_soundscape 1–4 sentences in one paragraph: ambience, physical sounds, non-verbal human sounds. Do not duplicate dialogue/diegetic music. Copy/reference relationships for the ambience/SFX layer go here: `The copied ambience layer from <Audio 1> continues throughout the target video.` ### 6. non_diegetic_music 1–3 sentences: instrumentation, tempo, dynamics; no abstract mood words. Audience-only copy/reference goes here: `<Audio 2> is directly reused as the complete audience-only score.` `N/A` if none. ## Hard constraints - Every label must map to a real input asset — invent nothing, drop nothing that was attached. - Duration is a hard constraint (4–15s): all timestamps strictly increase and fit inside it. - Six sections, brief text only; explanations separately and only on request. ## Mini example (first-frame anchor + voice timbre) ```text subject_definitions: <Subject 1> is the red-haired woman in <Picture 1>, with a freckled face, a cropped olive bomber jacket, and black cargo pants. <Picture 1> is the first frame of [Shot 1], showing the woman on a rooftop at dusk. <Audio 1> is the voice-timbre reference for <Subject 1> (S1), containing a low, steady female voice. summary: [keyframe completion + reference generation + audio reference] The target video opens on <Picture 1> and follows <Subject 1> as a gust of wind hits the rooftop; she speaks one line using the voice timbre referenced from <Audio 1>. retention_analysis: <Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - identity, red hair, freckles, olive bomber jacket, and black cargo pants retained. <Picture 1> ([Shot 1] first frame): fully_preserved - the shot begins exactly from the rooftop dusk composition. <Audio 1>: reference - only the low, steady timbre guides <Subject 1>'s delivery; the original signal is not copied. detailed_description: The target video is in a cinematic live-action style with cool dusk tones, vertical 9:16 framing. [Shot 1] The shot begins from <Picture 1>: <Subject 1>, the red-haired woman with a freckled face in a cropped olive bomber jacket and black cargo pants, stands at the rooftop edge. The camera arcs right with small amplitude at slow speed as a gust of wind snaps her jacket open and tears loose papers from a stack beside her into the air. She turns her head against the wind, hair whipping across her face, and <Subject 1> (S1) says in the low, steady voice referenced from <Audio 1>, <d>[English] There it is.</d> Her lips close and stay still. [Shot 2] At 00:05.000, the shot cuts to a tracking shot following one tumbling sheet of paper as it sails over the ledge into the blue dusk light; her fingers enter the frame a moment too late and close on empty air. The paper spirals down toward the street as the camera tilts down to follow it through the end of the video. overall_soundscape: A strong wind buffets the rooftop with flapping fabric and scattered paper rustle; distant traffic hums far below throughout. non_diegetic_music: A sparse synth pad at a slow tempo with a single repeating piano note, swelling slightly as the paper falls. ```