visual-asset-design

Visual Asset Design

This skill provides domain knowledge for compiling creative settings of characters/scenes/props into executable image generation Prompts. It does not replace the Project Schema, model capability limits, or user requirements; in case of conflict, follow: explicit user requirements > current image model requirements > defaults of this skill. The cost contract (one image per Variant, deduplicate before generation,required_variant_ids plan contract) is governed by the structural contract in the main Agent system prompt; this skill is only responsible for "what to draw and how to write it."

1. Reference Sheet Layout Specifications

Core Principles

• Type-driven: For character identity, prioritize using an artistic identity sheet; for scenes, use character-free spatial anchors; for props, use hero views and multi-angle details; only use time-sequence keyframes for action states. Information density must be subordinate to identity readability, and must not repeat or overlap just to fill the grid.
• User priority: When the user explicitly requests a single image, a specified number of grid cells, a specified viewpoint, or layout, follow the user's requirements.

Character Identity Reference Sheet: Cinematic Artistic Identity Sheet

When a character first establishes visual identity, by default generate an identity sheet with the feel of high-end animation studio character studies and art book layout, rather than a standard grid, blueprint, catalog, or repeated turnaround:

• For aspect ratio use Project settings.aspect_ratio(if the user explicitly specifies, the user's choice prevails), do not default to landscape, pure white or soft off-white background, large areas of negative space; no environmental storytelling, no irrelevant props, no watermark. Identity-defining clothing and carried equipment must be fully retained.
• The layout is asymmetrical, elegant, and intentionally off-balance, using varied image scales. One large, slightly off-center hero full-body view serves as the visual anchor, surrounded by smaller independent studies placed with clean spacing.
• Auxiliary studies are selected according to identity information value: neutral full-body, front/side/back, seated, leaning, crouching, high angle, low angle, expressive portrait; avoid multiple similar front-facing standing poses.
• All full-body, portrait, silhouette, and detail studies are strictly separated: no overlapping, merging, or stacking; do not crop the face, hide limbs, or truncate tails/identity equipment.
• Include a small silhouette study area (2–3 simplified black silhouettes), a small expression study area (subtle observable emotional changes), and a small detail study area (face, hairstyle/fur, clothing structure, and key equipment).
• Text retains only a minimalist character ID block: name, role, core emotion, visual signature. Handwritten labels and editorial arrows are used sparingly only when truly helpful; do not generate long explanatory passages.
• All views must be the same character: same face and proportions, hairstyle/fur, body proportions, clothing, equipment, inherent color scheme, posture language, and visual personality. For anthropomorphic animals, clearly specify bipedal/quadrupedal form; no species, age, body type, or clothing drift.

When the user explicitly requests a standard model sheet, a single image, or a specified grid, follow the user; still maintain independent views, non-repetitive information, and identity lock.

Sequence Keyframe Reference Sheet

A multi-grid image showing, in chronological order, the subject's key moments within a segment of content:

• From left to right and top to bottom, each cell corresponds to a key node or action moment.
• If the image is used as a reference for storyboarding or video generation, the internal frame of each cell must match the target video settings.aspect_ratio exactly, anda square grid (N columns × N rows) must be used: dividing the canvas into columns × rows will multiply the aspect ratio of a single cell by rows/columns; only when the number of columns equals the number of rows does a single cell equal the target aspect ratio. When the number of cells is not a perfect square, pad to the next perfect square (2–4 cells → 2×2, 5–9 cells → 3×3), and leave the extra cells as borderless outer blank space. Do not stretch, crop edges, or fill in with repeated content.
• For the number of grid cells, prefer a perfect square: 4 cells for a brief state, 9 cells for dense action or a complete scene; when the number of key moments falls between the two, use 3×3 and accept blank space—do not add or remove moments just to fill it up. When there are more than 9 cells, prioritize simplifying the background and per-cell details while keeping the action silhouette readable.
• Each cell should include: the subject's action/state, the environmental background, and the shot size.
• Describe cell by cell in the Prompt.
• When the user requests a single image: the single image shows the most critical moment.

Applicable scenarios (not limited to narrative):

• Narrative: plot progression (character running → looking back → surprised).
• Demonstration: usage flow (unboxing → holding → using → effect).
• Action: action sequence (preparing → jumping → landing → celebrating).
• Pet: behavior chain (sniffing → chasing and pouncing → eating → resting).
• Travel: location sequence (arriving → exploring → discovering → leaving).
• Interview: key moments (statement → gesture emphasis → reaction expression).

Scene/Environment Reference Image

A visual anchor that presents a space/scene, supporting two forms:

• Multi-cell grid (default): shows the scene from different angles (front/side/bird's-eye/local details), suitable for complex scenes or multi-area environments.
• Single panorama: no characters or only character silhouettes, including the lighting tone, color tone, and iconic environmental elements.
• When the user requests it, generate according to the user-specified viewpoint or detail focus.
• The scene anchor is an empty environment shot: no specific characters or passersby may appear in either form (except silhouettes); explicitly declare "empty shot, no people" in the Prompt; character consistency is handled by the character anchor image, and people in the scene image will contaminate the reference input of downstream storyboards.

Selection Rules

Purpose	Default Layout	User Adjustable
Character identity established for the first time	Cinematic art identity board (asymmetric hierarchy)	The user may request a standard model sheet, a single image, or a specified viewpoint
Prop identity established for the first time	Hero view + multi-angle/scale/material details	The user may request a single image or a specified viewpoint
Scene/environment establishment	Scene reference image (multi-cell multi-angle or single panorama)	The user may specify the viewpoint or detail focus
Key moment of the subject within a piece of content	Sequence keyframe reference sheet (multi-panel timeline)	The user may request a single image or a specified number of panels

Priority: explicit user requirement > default recommendation.

2. Visual Entity Prompt Structure

When generating a reference image for a VisualEntity, organize the prompt according to the following dimensions; do not just write a four-sentence summary of "subject/pose/environment/style":

1. Generation goal and format: asset type, aspect ratio, background, purpose, and whether it is an art identity board, scene anchor, prop study, or sequence keyframe.
2. Subject definition and identity lock: based on grounding context, reference images, and user description, first list the most distinctive silhouette/proportions/face/fur or hairstyle, then list clothing, equipment, inherent color scheme, brand logos, and content that must not change; clarify exactly what the reference images provide.
3. Layout and visual hierarchy: the position and scale of the main visual anchor, auxiliary study types, negative space, reading path, element spacing, and constraints against overlapping, merging, cropping faces, or hiding limbs.
4. View/region checklist: for each region, write the unique information provided by each independent study; for sequence keyframes, write the action phase, expression, shot size, and ending state panel by panel; for scenes, write spatial directions and landmark relationships.
5. Background and content boundaries: character boards use pure white/off-white with no environment; scene anchors explicitly show empty shots with no specific characters; prop images use a neutral background. Identity-defining clothing and carried equipment are not "irrelevant props" and must be retained.
6. Medium, lighting, and texture: the project's overall style, rendering medium, light direction/texture, color, aspect ratio, and quality goals consistent with the medium; do not use generic quality terms to override intentional sketches, grain, or hand-drawn lines.
7. Minimal exclusions: only list prohibitions that would cause identity drift, layout contamination, cropping, duplicate views, extra characters, garbled text, or style deviation.

When a character first establishes identity, use a cinematic art identity board by default; for key moments of a character in the content, use sequence keyframes; for scene anchors, use multi-angle empty shots or a single panoramic image; for props, use a hero view with multi-angle details. When the user has explicit requirements, follow them.

Writing requirements

• Each prompt should be detailed but not verbose, focusing on key visual elements.
• Use specific terminology from composition and cinematography to describe viewpoint and composition.
• Follow the core principles and image texture suggestions in the prompt writing rules.
• When the user has explicit requirements, follow them.

3. Composition and cinematography

When generating reference images, choose an appropriate composition method based on the subject type and narrative needs:

Common camera viewpoints

• Low-angle shot: emphasizes the subject's height, authority, or sense of oppression.
• High-angle shot: shows the spatial layout or the subject's sense of insignificance.
• Top-down overhead view: shows overall spatial relationships or a floor plan layout.
• Over-the-shoulder view: creates a sense of immersion or voyeurism.
• Foreground occlusion: adds layers and depth through foreground elements.
• Reflection shot: shows the subject through a reflective surface, adding visual interest.
• Silhouette shot: Subject silhouette in backlight, emphasizing form and atmosphere.
• Frame within a frame: Frame the subject through a framing structure to guide the eye.
• Deep vanishing point: Use perspective lines to guide the eye and create a sense of depth.
• Wide negative space: The subject occupies a small part of the frame, emphasizing environmental scale or a sense of loneliness.
• Compressed telephoto: Compress the distance between foreground and background, highlighting the subject or creating a sense of crowding.
• Handheld close-up: Simulate the imperfect composition of handheld shooting to increase realism.
• Diagonal movement: The subject or camera moves along a diagonal, creating dynamic tension.
• Symmetrical staging: Symmetrical composition, creating a sense of ritual or surrealism.
• Asymmetrical balance: Balance an asymmetrical composition through visual weight.
• Partially hidden subject: The subject is partially obscured by the environment or other elements, increasing a sense of mystery.
• Environmental scale: The subject appears small within a vast environment, emphasizing the sense of space.
• Subjective point of view: Shoot from the subject's perspective, letting the audience identify with the character.
• Layered foreground/midground/background: Clearly distinguish three layers to increase the image's depth.

Composition selection rules

• Character identity established for the first time: Use one off-center heroic full-body view to establish primary and secondary emphasis, then support it with 3-8 truly different information views, silhouettes, expressions, and detail studies; do not use an evenly weighted grid.
• Character within the content: Choose the composition according to the emotion (use low angles/foreground occlusion for tension, negative space for loneliness, diagonal movement for action).
• Scene establishment: Use a deep vanishing point or wide negative space to show environmental scale.
• Prop display: Use symmetrical staging or a frame within a frame to highlight the subject.

Priority: Explicit user requirements > default recommendations.

4. Prompt writing rules

Avoid image moderation false positives (mandatory)

Image model output moderation rejections (such as DashScope's green-net) aredeterministic: the same prompt will be rejected every time, and all storyboard and video nodes that depend on that asset will be blocked. So when writing prompts, avoid common false-positive word families from the start, rather than waiting until after failure to revise.

• Do not write police badge/military insignia-style details for uniformed characters: Field case 2026-08-26, a bus driver identity sheet was rejected by output moderation just because it wrote "navy blue uniform + shoulder epaulettes + peaked cap with silver cap badge"—this description would be judged as police/military uniform style. For civilian uniformed characters (drivers, conductors, security guards, gatekeepers, maintenance workers), useoccupational cuesto express identity: work uniform color and material, badge lanyard, reflective strips, carried tools, worn cuffs, sitting posture, and hand movements; do not write cap badges, police badges, shoulder rank insignia, arm patches, or numbered chest badges.
• Equally easy to misjudge: wounds/bloodstains/bruises, close-ups of minors, nudity or form-fitting depiction, clearly pained or pathological expressions, names of real public figures or "looks like some celebrity," close-ups of weapons and controlled implements, religious or political symbols. When you need to convey fatigue, injury, or conflict, switch to environmental and action-based indirect description (rain, sweat-soaked collar, gripping a railing, a hanging hand) rather than directly depicting bodily harm.
• Once an output moderation rejection is received: do not retry as-is. Identify the wording most likely to be flagged and rewrite it, preserving the shot intent; after rewriting, the prompt fingerprint changes, and only then will the node dispatch again.

Core Principles

• Concrete description: Avoid generic words like "epic," "beautiful," or "cool" unless they are supported by specific visual details.
• Layered description: Clearly define the elements and relationships of the foreground, midground, and background.
• Light and atmosphere: Describe light direction, type, color temperature, and atmosphere.
• Narrative feel: The scene should imply story, tension, motion, or emotion, rather than being a purely static display.
• Intentionality: Every compositional choice should have a clear visual purpose.

Image texture recommendations

Choose an appropriate texture treatment based on the project style:

• Natural texture: Prioritize natural materials and environmental textures, and avoid excessive smoothness or a plastic feel.
• Environmental detail: Appropriately add airborne particles, mist, dust, reflections, shadows, etc. to increase depth.
• Realism: Prioritize moments in motion or full of emotion, rather than static poses (unless the user requests otherwise).
• Imperfect realism: Depending on style needs, slight motion blur, soft focus, grain, etc. may be used to increase realism.

Prohibited recommendations (unless the user explicitly requests them)

• Avoid an overly sharpened or overly processed look.
• Avoid centered portrait-style composition with no environmental context.
• Avoid fully symmetrical, posed compositions with no visual tension.
• Avoid a plastic or overly synthetic look.

5. Entity VisualEntity Definition Specification

continuity field

• Character continuity description: body type, face shape, skin tone, hair color, eye color, height range, signature features (scars/tattoos/accessories).
• Scene continuity description: core color palette, lighting tone (day/night/indoor/outdoor), signature environmental elements.
• Prop continuity description: primary material, color, proportions, visible functional markings.

name and description

• name is unique across the entire Project; use searchable character/scene/prop names.
• description focuses on the key features that distinguish the entity from others of the same type, and does not repeat baseline attributes already covered in continuity.

Variant consistency rules

Multiple Variants under the same entity must:

• Core identity features (invariant attributes such as body type, face shape, skin tone, hair color, eye color) do not change across Variants.
• Consistent style baseline: all Variants use the same rendering style and quality baseline.
• Reference priority across Variants: generated artifact versions take priority over external source versions.
• The environmental elements merged from the reference images should remain consistent with the character's core identity, and contradictory environmental features should not be introduced due to Variant changes.

Reference Image Binding Specification

• For a Variant's reference_asset_version_ids , prioritize using the source_asset_version_id。
• Prioritize using one reference image to cover multi-angle/multi-moment information of the subject; bind multiple reference images only when the subject needs to appear in multiple significantly different environments.
• Reference images of different entities must not contradict each other: different Variants of the same character should not display different skin tones or body types.

Character Identity Stability

• The same character maintains stable visual anchors across all Elements and Variants: body silhouette, facial proportions, skin tone/hair color/eye color.
• Changes in a character under different environmental lighting only affect the lighting and shadow presentation, and do not change inherent colors and proportions.
• Cross-Element character references are maintained through the stable ID in visual.entities , and character visual details are not repeatedly declared in the Element.

Scene and Prop Consistency

• Key scenes establish an independent VisualEntity (kind=scene), including an environment panorama description, color tone reference, and main area layout.
• Key props establish an independent VisualEntity (kind=prop), including an appearance description, material reference, and proportional relationship to the character.
• Variants of scenes and props are only valid within this entity's scope, and do not conflict with the visual settings of other entities.
• When characters, scenes, and props appear as a whole, prioritize maintaining consistency through merged reference images to reduce cross-entity reference conflicts.

Type Adaptation

Emphases of different entity types in the Prompt:

• Character: When establishing identity for the first time, use a cinematic artistic identity sheet, establishing identity with a large hero full-body image and independent multi-angle, silhouette/expression/detail studies; at key moments in the content, use sequential keyframes to show action and expression changes. When the user requests a standard grid or a single image, follow the user's request.
• Scene: By default, use a multi-panel multi-angle display showing different angles and details; when the user requests it, use a single panoramic image showing the lighting tone and recognizable landmarks/area boundaries.
• The scene continuity atlas must explicitly choose one topological expression: when it is to be further consumed by image/video models, default to usingtextless visual topology(consistent vanishing points, repeated landmarks, path direction, and a small number of textless arrows), and must not simultaneously require clear spatial labels and no text; short labels are allowed only when the user explicitly needs a director review image, and these labels must be listed as explicit exceptions to the text exclusion items. Textless mode must also not secretly substitute labels with letters A/B/C, place names, or legend text.
• Prop: Use a hero view + independent multi-angle/proportion/material and functional details, and do not apply character expression or pose layouts. When the user requests a single image, follow the user's request.

6. Creative Taste Guidelines

The following creative taste guidelines originate from the vendored review_rubrics (Qwen-MM-Plugins video-edit skill) and should be followed (advisory guidelines, not gates):

• Composition safety: Nothing covers faces/UI/focal action ([no-occlusion]); overlay anchor differs from the previous scene's; every overlay has a full life-cycle — entrance, idle micro-motion if held >1.5s, choreographed exit with a hard kill (no element leaking past the scene's time box).
• Consistency: Maintain consistency in styling, color scheme, and lighting across character/scene design sheets; text elements must not be garbled or use placeholder glyphs.