0. Goals and scope of application
Make cinematic feel come from camera position, the relationship between characters and space, visible action, practical light sources, and editing between shots, rather than adjectives, filters, or piled-on grain.
Only default to live-action cinematic texture when the user has not specified a medium and has not asked to preserve a reference style. When hand-drawn, animation, product, or commercial space is specified, comply with the corresponding medium and asset requirements, and do not judge "non-live-action" or "advertising project" itself as failure. Add titles, posters, and visual systems only when the user requests them.
Processing order: user-given plot and asset facts -> continuity and spatial feasibility -> shots and action -> aesthetics. The following protocol takes precedence over fixed beats, composition quotas, reference image isolation, and medium restrictions in old style references.
1. Determine the deliverable first; do not automatically expand the task
| Intent | Deliverable |
|---|---|
| Single cinematic still | 1 independent shot |
| Three-panel, default cinematic test | 3 independent shots, then stitched vertically |
| Nine-shot, nine images telling a story | 9 independent shots, then externally assembled into 3x3 |
| Storyboard video prompts, camera movement plan | Director cards and copy-ready short prompts corresponding to the number of shots; do not automatically generate images or video |
| Storyboard images with accompanying video prompts | Independent images and corresponding director cards; clearly state whether the image is the first frame, midpoint, or last frame |
| Actually generate video | First verify available video tools and capabilities, and execute only within the scope authorized by the user |
The quantity is based on the user's request. If no number of stills is specified, default to three-panel; do not arbitrarily expand to nine shots just because it is being converted to video.
Default horizontal 2.39:1; the strict 21:9 ratio is 7:3, not 2.39:1. When the user specifies an exact ratio, preserve that ratio and check the generation and layout dimensions; do not stretch. Adapt the video aspect ratio to the selected backend's capabilities; cropping must preserve the narratively key areas.
Comply with the backend specified by the user. When the official built-in Image Gen is specified, do not silently switch to Maliang or another service; an image tool is not a video generator. An unknown backend does not prevent writing neutral video prompts, but do not fabricate API parameters or unsupported control features.
2. Plot and reference characters
For multi-shot stories, first read story-state-continuity.md. For nine shots, additionally read nine-shot-story-protocol-v3.md。
- Complete story or already specified shots: by default faithfully storyboard, preserving events, order, relationships, and ending.
- One-sentence open premise: original creation is allowed; distinguish user facts from newly added design.
- Expanding three-panel to nine shots: prioritize adding action stages, approach processes, sightlines, and environment; do not force in reversals, sacrifices, or open endings.
- Mark the purpose of each input: inspiration / style / character / scene / product / plot anchor frame; multiple characters are allowed, but clearly state the scope to preserve.
- The rule of "extracting only inspiration" cannot override assets the user asks to restore. Approved plot anchor frames cannot be rewritten by style templates.
When preserving a reference style, an existing calibration frame, or correcting style drift is required, read style-reference-fidelity.md. Distinguish "reference not used as a single frame of the final film" from "reference must not be input into generation"; do not misread the former as the latter.
For assets, new camera positions, or spatial continuity tasks, read asset-space-protocol.md. Make unknown spaces explicitly unknown; do not disguise speculation as three-dimensional measurement.
3. Cinematic decisions
For each shot, determine one main event, a viewing standpoint, and an information landing point; micro-actions, pauses, and observation can all work, and conflict need not be manufactured. First ask what the audience should know at this moment, then decide camera position, shot size, and occlusion.
When the user explicitly restricts shot size (e.g., "no close-ups; focus on the scene and action relationships"), write it into the shot ledger; do not insert close-ups of eyes, hands, or objects that violate the restriction for the sake of refinement or shot variety.
- Composition arises from relationships: distance, intimacy, obstruction, waiting, power, or observation, rather than applying a template of "wide shot—subjective shot—eye close-up."
- Gaze flow: where it enters, what guides it, and what key evidence it lands on. Not every shot must be complex.
- Color comes from clothing, architecture, weather, and light sources. First lock the overall color and key light logic; do not force every story to have a fixed number of colors.
- Optics and materials: credible exposure, natural skin, localized softening, moderate grain. Avoid unmotivated HDR, plastic skin, neon, particles, and global chromatic dispersion.
- Arrange the rhythm of shot size and information density across the whole segment, retaining necessary shot-reverse-shot, repeated compositions, and ordinary transition shots. Do not force every adjacent shot to change four items each time.
- Light sources remain continuous in their position within the world; changing camera position can change the direction of illumination on screen, which does not mean the real light position moved.
4. Continuity: separate identity from state
Establish a lightweight Continuity Bible, repeating only the facts relevant to this shot:
世界：时代、时间、地点、可见结构、光源逻辑、媒介
身份：角色 ID、少量稳定外貌/服装锚点、关系
实体：关键道具 ID、数量、外观；不要把可转移物件当永久身份锚点
状态：每镜的持有人、位置、开合/损坏、动作阶段、可见性；化形故事另记意象/半成形/可辨实体
证据：计划状态与实际图像观察分开记录
After a camera, wristband, umbrella, etc. is handed over, the original holder cannot automatically carry another one merely due to "identity continuity." Occluded/off-screen does not equal nonexistent; a reflection does not equal a newly added entity. Returns, repairs, and flashbacks permitted by the story require clear event-based justification.
Before generation, establish a shot ledger:
镜号 | 来源节拍 | 功能 | 镜前→主要事件→镜后 | 机位与轴线 | 可见证据 | 依赖镜号
A still frame represents only one stage of an action; only states that have actually passed acceptance enter downstream reference. Do not declare that an action has been successfully generated based on a prompt description.
5. Space and reference input
Divide scene assets into "visible facts / grounded speculation / unknown areas." Separate world positions, character action axes, and screen left-right.
For each shot, determine: the camera's allowable standing area, shooting direction, shot size, necessary anchors, occlusion, action reachability, and the unknown spaces that would be exposed. When a single image has no credible reverse side, first use a conservative camera position; a major reverse angle or orbit requires additional views, or a spatial extension explicitly accepted by the user.
Check character proportions in combination with pose, depth, ground contact, and same-depth references, not with fixed pixel sizes. Fixed windows, doors, and products must not be duplicated or moved due to a change in camera position.
Select reference inputs by role: style reference, identity reference, spatial reference, current state reference. A small number of necessary references is better than indiscriminately inputting all of them. Original references can support style/identity/space, but cannot override state changes that have already occurred later; nor can they reference only the previous image and accumulate errors.
6. Still-frame prompts
Generate each shot independently; do not let the model directly draw a triptych, nine-grid, or storyboard sheet. After generation, composite externally, keeping source images that can be replaced individually.
Image prompt order:
- Single frame, medium, and explicit aspect ratio.
- Identity, space, and current state related to this shot.
- Action stages and key evidence for this shot.
- Camera physical position, direction of shooting, shot size, and perspective intent.
- Decisive foreground/midground/background information.
- Light sources, overall color, and limited imaging characteristics.
- Condensed exclusions compatible with the backend.
Use live-action base as needed, do not stack all at once:
standalone live-action film still, practical location, natural skin texture, soft highlight roll-off, restrained production design, subtle uneven grain
Do not copy still-image negative terms, long character descriptions, or "focal length tags" verbatim into image-to-video. Focal length text expresses intent and does not prove the backend precisely executed optical parameters.
7. Video action and camera movement
When the user wants video prompts or actual video, you must read video-motion-protocol.mdwhen writing examples are needed. video-prompt-examples.md。
For each shot, deliver a "Chinese director control card + copyable short prompt", with language following the user's requirements. The director card manages the first and last states, timing, movement, and transitions; the short prompt contains only the visible action and camera behavior needed for execution, and does not copy the entire card.
Key constraints:
- First determine whether the input image is the action's first frame, midpoint, end, or a composition reference, then decide where to start the motion.
- Character movement, camera displacement, rotation, focal length changes, and focus pulls are described separately and must not be substituted for one another.
- By default, one segment has one main event and one main camera movement; a fixed shot also counts. Complex long takes are designed only when needed.
- Clearly specify the reference object, direction, ending composition, and stopping condition of the camera movement, avoiding abstract instructions like "advanced camera movement".
- Nine shots are usually nine independent segments; a cut is not a camera moving through a wall, and the first frame of the next shot does not automatically become the last frame of the previous shot.
- Design duration and actual generation duration are separate; when there is no corresponding control interface, do not promise execution precise to meters, angles, frames, or seconds.
- If no video has been generated, it can only be marked as "prompt checked, motion unverified".
8. Orchestration, rework, and records
The queue may be organized in three shots, but first accept the key state-change shot, then generate the shots that depend on it. Do not mechanically use the first initial-state image to guide all subsequent shots.
For each record, note the shot number, revision, exact prompt, input reference and role, return path, planned and observed state. If the tool provides model version/seed, record that too; do not fabricate. Order cannot be inferred from modification time alone.
When it fails, locate the cause: contradictory instructions / backend noncompliance / outdated reference state / insufficient capability. Fix only the corresponding shot and recheck downstream dependencies. By default, at most two automatic targeted reruns per shot; if it still fails, report the failed items and required supplements, and do not retry indefinitely. If the user sets another budget, follow it.
When encountering a safety block, first check whether the request is allowed; do not bypass restrictions by hiding core facts. Only attempt a compliant rewrite when the original request itself is allowed and the alternative expression remains faithful; if changing the core plot is required, explain and ask.
The nine-shot composite uses the existing compose-nine-shot-storyboard.ps1, first verify the parameters and target ratio. Preserve 9 source images, 3 vertical triptychs, and 1 contact sheet; default to small black gaps, no additional text, and no stretching. Output to the directory specified by the user, or if none is specified, use a task-specific directory.
The accompanying script uses PowerShell and System.Drawing, and is preferably executed on Windows; in other environments, use available equivalent tools to complete the same composite, without replacing source images that have already succeeded.
9. Acceptance: facts first, then art
Hard gates: key facts of the original story, identities and mandatory characters, item counts/ownership, specified art style/medium, spatial topology, action reachability, input first-frame compatibility, file integrity. When the user requires recognizable characters, a shadow or smoke silhouette alone does not count as complete; embodied identity and origin are verified separately.
Results are recorded as "pass / fail / cannot determine," with visible evidence attached; if it cannot be seen clearly, it cannot count as a pass. A hard gate failure must not be offset by an overall aesthetic score, nor may it be claimed as "resolved." When delivering a draft, clearly mark the items that did not pass.
- Across shots, check both adjacent continuity and long-distance causality; for example, after shot 5 leaves a wristband behind, the wrist in shot 9 should still be bare.
- For restoration checks, use an "original story fact -> shot evidence" mapping; when necessary, retell the story without looking at the notes. Do not call a self-check in which the answers were seen a blind test.
- For video checks, examine the beginning, middle, end, and key moments of action, distinguishing subject motion from camera motion; a still frame cannot prove temporal facts such as "half a beat late."
- After the facts pass, evaluate composition, color, rhythm, material, and medium; do not impose the same shot-count quota on different subject matters.
- When only checking prompts/rules, do not report image or video success rates.
10. On-demand resources and delivery
- Retain reference art style, calibration frames, and style correction:style-reference-fidelity.md。
- Files, prompts, and packaged delivery:delivery-contract.md, read when preparing delivery files.
- Multi-shot narrative, item changes:story-state-continuity.md。
- Nine-shot organization:nine-shot-story-protocol-v3.md, retain the old file name for compatibility with existing references.
- Assets, changing angles, proportions, spatial axes:asset-space-protocol.md。
- Video prompts/videos:video-motion-protocol.md。
- Camera movement phrasing examples:video-prompt-examples.md, not a template library loaded every time.
- Image texture needs adjustment:cinema-dna-v4-anti-ai.md, v4 is a historical style patch name, not the current skill version.
- When optics, subject matter, or poster methods need to be expanded, read by section cinema-dna-full-spec.md, not loaded in full by default.
Image only: deliver the image and necessary verification notes. Prompt only: deliver per-shot prompts, without automatically calling the media backend. Image plus prompt: specify shot number, image path, frame role, corresponding prompt, and unverified items. Do not read the sample image library by default or append a poster.
For file delivery, use directly copyable TXT prompts by default, with the directory language following the user; use Chinese directories for Chinese projects. The language and format specified by the user take priority; do not replace a TXT or other format they explicitly requested with an MD file.