StarAI-2026/ComfyUI-MiniMax-H3-SkillBridge

High-Dynamic Camera Movement (H3 Image-to-Video Prompt)

Transforms reference images automatically loaded by the plugin into a complete MiniMax H3 image-to-video Ref2VA prompt. This skill only produces video prompts; it does not generate character cards, action first-frame text prompts, or storyboard tables.

Input Modes

• Image Mode (Recommended): The plugin has automatically loaded 1-6 reference images. Use the tag rules from h3-ref2va-contract.md to define each image: <Subject N> defines reusable elements (character, scene, prop, style), <Picture N> is used only when the image serves as a first-frame/keyframe anchor.
• Pure Text Mode: If no reference images are provided, do not error or refuse. Instead, create subject definitions in subject_definitions based on the user's text description, e.g., <Subject 1> is a ... described by the user as .... Do not introduce <Picture N> first-frame anchors; use [reference generation] for the summary; only if the user explicitly provides an action composition description and intends it as a starting frame should that description be declared as a starting frame anchor with keyframe completion.
• Neither mode outputs character cards or action first-frame text prompts. Image generation tasks are outside the scope of this skill.

Reference Files

• references/h3-ref2va-contract.md: The single authoritative contract for the six-segment Ref2VA structure, tags, reserved markers, timeline, and sound rules.
• references/high-motion-h3.md: High-dynamic definitions, camera command syntax, motion density and causal chains, rejection checklist.
• references/generalized-speed-grammar.md: Motion prototype library, camera module library, flexible time allocation, and high-speed vocabulary.

Do not read visual-case-design.md or sample-motion-blueprint.md. Unless the user explicitly requests to replicate the hoverboard sample structure, do not apply its fixed timepoints, jump, 180-degree orbit, or landing drift.

Output Format

Strictly output two sections in node format:

[Visual Analysis]
Summarize the subject, scene, action, and available style anchors from the reference images or text description in 2-3 Chinese sentences.

[Final Video Prompt]
English six-segment format, with field names and order strictly fixed:
subject_definitions:
...
summary:
...
retention_analysis:
...
detailed_description:
...
overall_soundscape:
...
non_diegetic_music:
...

detailed_description must fully cover the composition, subject identity, environment and lighting, action and state changes, camera movement, synchronized physical sound effects, and where reference content takes effect. Do not fabricate actions for word count; time feasibility takes priority.

High-Dynamic Hard Requirements

Every prompt must simultaneously satisfy:

1. The subject displaces along at least two axes: forward + descent, lateral + rise, forward + lateral rotation/lean, etc.
2. The subject's apparent scale or camera distance changes significantly: filling the frame, pulling back, passing close to the lens, rapidly changing distance.
3. Foreground structures, cables, railings, debris, or particles cut across the frame, creating rapid parallax.
4. The subject interacts with physical elements: passing through a door, crossing an obstacle, scraping a rail, weaving around a vehicle, hitting a surface, deflecting thrust, overtaking a vehicle, brushing past.
5. Acceleration/deceleration is visible: body compression, thrust shape, wake distortion, sparks, dust, water separation, shockwaves, clothing/hair lag.
6. The camera actively tracks: push in, pull out, lateral movement, pedestal up/down, tilt, change distance, follow a lean – do not remain centered and static or at constant speed parallel tracking.

Write each beat as a causal chain: subject force/action → body/equipment response → environmental reaction → camera response. Use speed contrast; prohibit constant-speed parallel tracking throughout.

Camera Command Syntax

• Official bracket commands: [Truck left], [Truck right], [Pan left], [Pan right], [Push in], [Pull out], [Pedestal up], [Pedestal down], [Tilt up], [Tilt down], [Zoom in], [Zoom out], [Shake], [Tracking shot], [Static shot].
• Write simultaneous commands within the same bracket separated by commas, e.g., [Tracking shot,Push in]; a single bracket may contain at most 3 commands; use different command groups at different time points.
• High-dynamic mode prohibits [Static shot] and forbids simultaneous opposing commands such as push in + pull out.
• Use natural English descriptions for arcs, orbits, crane arms, flips, close passes, and other movements without official commands; do not fabricate nonexistent bracket commands.
• Arrange at least 2-3 sequential camera command groups covering tracking, direction change, and conclusion.

Time Allocation

Allocate beats based on video duration, without fixed timestamps:

• Approximately 5 seconds: 3 beats, 0.0-0.5 immediately enter high speed, a composite crossing or direction change in the middle, a decisive spatial closure at the end.
• 6-8 seconds: 3-4 beats, allowing one extra reversal, obstacle, or environment reveal.
• Approximately 10 seconds: 4-5 beats, including one major speed contrast and one environment transition.

Timeline and Sounds

• Do not write At to cut time at the beginning of [Shot 1]; for subsequent shots, use At MM:SS.mmm strictly increasing and not exceeding the total duration.
• If using <Picture 1> as a starting anchor, explicitly write the shot begins from <Picture 1>.
• Use <d>[language] ...</d> for dialogue/lyrics and retain the original language; identify speakers consistently with (S1), (S2).
• Place synchronized dialogue, vocals, and in-shot sound events in detailed_description; place continuous environment and physical action sounds in overall_soundscape; place only audience-audible music in non_diegetic_music, specifying instruments, tempo, rhythm, and dynamics; write N/A if missing. The three sections must not repeat each other.

Self-Checklist (Pass Each Item Before Output)

• Each of the six fields appears exactly once in strict order.
• All tags are defined in subject_definitions before use and maintain consistent meaning throughout.
• The summary task type matches the input mode: pure text reference uses [reference generation]; image with first-frame anchor uses [keyframe completion + reference generation].
• Each tag appears on exactly one line in retention_analysis, using fixed markers (fully_preserved, partially_preserved, attribute_transfer, weak_reference; audio uses fully_copy, partially_copy, reference, weak_reference).
• detailed_description opens with [Shot 1], [Shot 1] has no At timestamp, subsequent shots have increasing timestamps within the duration.
• At least two sequential official camera command groups; no [Static shot]; no bracket exceeds 3 commands.
• Satisfies the six high-dynamic hard requirements: two-axis displacement, scale change, foreground parallax, obstacle contact, visible acceleration, active camera.
• Causal chain is complete: action → body/equipment response → environmental reaction → camera response.
• At least 3 explicit timepoints covering start, crossing, and closure.
• overall_soundscape and non_diegetic_music do not repeat dialogue and do not repeat each other.
• The ending can be achieved from the starting state within the given duration; do not use static gestures or blinking as closure.
• Image mode only uses reference images provided by the plugin; pure text mode creates subjects based on user description, without fabricating undescribed details.