SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsvisual-identity-expander

visual-identity-expander

General & other

View repository

Updated 10-15 21:40

About this skill

Visual Identity Expander

Expand one core reference image into a set of visual identity solutions with a unified visual language that can continue to be iterated on.

Core principles:First lock down the recognizable visual genes, then extend to usage scenarios. Each image can vary, but it must be evident that they belong to the same person, character, brand, or product.

Define the deliverable first: images, or prompts

This determines how the identity is anchored; clarify this first before proceeding.

  • Deliver prompts(default): the identity must betextualized—write the reference image's features as a self-contained English visual description. The prompt cannot survive by relying on "refer to the uploaded image," otherwise it is worthless once it leaves this session.
  • Deliver images: the reference image is in the generation context, so an image anchor can additionally be attached Image 1: identity reference, for higher fidelity.

Default dual anchor: write out the text anchor fully, and when an image is available, attach one line of image anchor as well. The same prompt maintains fidelity when accompanied by the image, and can still be used in a degraded form without the image. For rules, see references/prompt-contract.md。

For people, state the boundaries truthfully: text can lock down hairstyle, glasses, clothing, temperament, and camera habits, but cannot lock down the same face. When delivering pure prompts, state this at the top of the output.

Read

  • Extract and judge visual genes: read references/visual-dna.md。
  • Compile prompts: read references/prompt-contract.md(dual anchor + seven slots + grouped base).
  • Choose the deliverable form and decide how many to produce: read references/deliverables.md。
  • When a complete walkthrough example is needed, read references/examples.md。

Underlying writing rules inherit gpt-image-prompt-spec 's seven slots and five hard rules; this skill does not redefine them, and only supplements the visual identity extension layer.

Input

Prioritize inferring from user input and uploaded images; do not force form-filling:

  1. Visual matrix: Logo, portrait avatar, personal image, IP, illustration, product, or object.
  2. Name and text: brand name, person's name, account name, slogan; if absent, do not fabricate.
  3. Purpose: personal brand, commercial brand, content account, product promotion, event, or portfolio.
  4. Temperament: infer from the reference image; when the user explicitly specifies it, follow the user's requirement.
  5. Aspect ratio: determine by platform and purpose; when the user specifies a ratio, follow it strictly.

Ask questions only when missing information would significantly change the result. For everything else, make reasonable assumptions directly and briefly explain them at the top of the output.

Workflow

  1. Define the deliverable: whether to produce images or prompts determines the anchor strategy.
  2. Identify the input type: Logo, person/avatar, IP/illustration, product/object, or a combination of multiple inputs.
  3. Extract visual DNA: break it down into immutable features and variable features, writing each as a compilable visual fact—cover the original image and read it aloud; someone who has never seen the original image should be able to draw it. Techniques are solidified into identity; content decisions in the reference image should be replaced.
  4. Establish system rules: primary and secondary colors (give color values), shape language, materials, lighting, typography, camera, and mood.
  5. Determine scale: check deliverables.md for the item counts based on what the user actually wants, and do not force five items. The complete set is — key visual / identity system board / variation system / application system / communication scenarios.
  6. Compile prompts: output each item in seven slots, complete, independent, and ready to use directly. Group base (identity lock, color palette materials, identity discipline)must be written verbatim into every item, with no undecided options, slash menus, or blank placeholders left.
  7. Consistency check: verify item by item whether identity, silhouette, color, material, text, and visual temperament have drifted; if the base wording branches, go back and unify it.
  8. Generate images as needed: when the environment provides an image generation tool and the user explicitly requests it, generate them one by one; otherwise deliver only the prompts.

Adaptive Routing

  • Logo: preserve geometric proportions, negative space, wordmark relationships, and brand colors; the variation system changes to auxiliary graphics, patterns, and Logo variations, and does not generate a character IP by default.
  • Portrait avatar/personal image: preserve facial identity, hairstyle, glasses, signature clothing, and temperament; the variation system generates actions and expressions.
  • IP/illustrated character: preserve silhouette, facial feature proportions, color scheme, and signature props; the variation system generates a character design and pose sheet.
  • Product/object: preserve shape, material, structure, and brand marks; the variation system generates viewpoint, detail, and theme variations.
  • Multiple reference images: first distinguish the primary reference from auxiliary references. The primary reference locks identity; auxiliary references only provide style, material, or scene, and do not blend into a new identity.

Output Format

## 输入判断
- 输入类型：
- 交付物：提示词 / 图
- 锚点策略：双锚 / 纯文字 / 带图
- 使用目标：
- 必须保留：
- 可以变化：
- 本次假设：

## 视觉 DNA
- 核心轮廓：
- 主辅色（带色值）：
- 字体与图形语言：
- 材质与光线：
- 气质关键词：

## 01 主视觉
```text
完整提示词
```

（按定好的规模继续 02–05，格式相同）

## 一致性检查
- 身份：通过/需修正
- 轮廓：通过/需修正
- 色彩：通过/需修正
- 文字：通过/需修正
- 气质：通过/需修正

Quality Gates

  • Each prompt must be usable independently, without relying on "same as above" or omitted information from earlier text.
  • Repeat the group base verbatim: the three sections identity lock, color palette materials, and identity discipline must be identical word for word in every item. Any branching wording is a compilation error.
  • Do not rely on "reference uploaded image" to carry identity— when the deliverable is a prompt, the identity must already be textualized to the point of being self-sufficient.
  • Write exclusions into the Constraints body,do not output a separate Negative prompt: block。
  • Do not use filler words like masterpiece / 8k / ultra detailed; replace them with visual facts.
  • Do not treat "looks good" as consistency; check identity and system rules first, then check aesthetics.
  • Do not fabricate text for Logos, packaging, or clothing that cannot be confirmed. When accurate text is needed, explicitly write the original text provided by the user, and add render all text verbatim。
  • When color values cannot be given accurately, give the color family plus a tendency; do not make up hexadecimal values.
  • Character extensions must not arbitrarily change face shape, age, ethnicity, hairstyle, and signature accessories.
  • IP extensions must not arbitrarily change silhouette, facial feature proportions, color scheme, and signature props.
  • Product extensions must not arbitrarily change structure, interfaces, label hierarchy, and core materials.
  • Communication scenarios need real ambient light; writing studio lighting sends it back to a display board.
  • Do not fabricate already-generated results. When the image generation tool has not been called, state only that the prompts have been completed.
  • Avoid multiple images turning into several unrelated style cases; the variation must occur within a unified rule. Also avoid the opposite—each entry repeating the original scene of the reference image.