Home / All Skills / Design & Motion / rare-style-explorer
English (EN)简体中文 (ZH)日本語 (JA)
Design & Motion · vibeshotclub/vsc-skills

rare-style-explorer

Generate and refine AIGC image prompts by combining rare, prompt-ready visual sub-style tags from a bundled 620-entry style library. Use when the user wants style exploration, image prompt variants, rare visual styles, non-generic aesthetics, style mixing, prompt matrices, or subject-to-style ideation for image generation.

L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandDual-pass Audit
Author vibeshotclubGitHub vibeshotclub/vsc-skills ↗Stars 832License 仓库未提供 LICENSE 文件（find 全仓无 LICENSE*；GitHub API license 字段为 null）Commit fcee81beb6
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Capabilities are powered by a structured library of 620 styles plus a deterministic, reproducible weighted random combinator; SKILL.md solely governs invocation conventions and output discipline.

rare-style-explorer/SKILL.md
The bundled library lives at `references/style_library.json`. Load it only when the user asks for specific style lookup, filtering, auditing, or manual curation. For normal prompt generation, run `scripts/explore_styles.py`.
Actual JSON read: The styles array has a length of 620, consistent with the 620 claimed in SKILL.md, README, and the description.

The library consists of 'keyword-first' structured entries: each entry includes Chinese and English names, category, visual DNA, combination role, pitfall risks, and mitigation tips, ready to be directly assembled into prompts.

rare-style-explorer/references/style_library.json
"中文风格名": "液态铬金属",
A single entry includes "English prompt tokens": "liquid chrome, molten reflective metal", "category": "Materials and Surface Textures", "pitfalls", "mitigation tips", and "example phrases".

The combination grammar relies on fixed slots: Subject + one strong primary style + zero or one surface/light layer + zero or one format/space layer + anti-drift negative constraints.

rare-style-explorer/SKILL.md
Use one strong base style. Add zero or one surface/light style. Add zero or one format/space style. Add zero or one defect layer only when a more analog or media-specific finish is useful.
SKILL.md also provides text templates such as {subject}, {base_style}, {surface_or_light}, {format_or_space}.

The script applies templates based on modes rather than generating freely: each of the six modes has its own sentence skeleton.

rare-style-explorer/scripts/explore_styles.py
"minimal": "{subject}，{base}，{support}，主体轮廓清晰，视觉识别度强，高细节，{negative}",
TEMPLATES also includes five options: product, character, poster, scene, and material-series.

Generic word suppression + specificity weighting: entries identified with an excessively high ratio of generic words in their tokens have their weights reduced to 0.25.

rare-style-explorer/scripts/explore_styles.py
if len(terms) <= 1 or specificity < 0.35:
The GENERIC_WORDS set includes cinematic, retro, surreal, vintage, etc.; style_weight() boosts entries containing a 'source batch' by a factor of 3.0 when freshness=high.

In-batch deduplication relies on token overlap: entries sharing ≥2 content words and accounting for ≥50% overlap with already selected entries are judged as overly similar.

rare-style-explorer/scripts/explore_styles.py
def too_similar(style, seen_term_sets):
pick_unique() applies a 4-level relaxation mechanism (first loosening categories, then similarity, and finally used status) to ensure enough items can still be gathered when count is large.

Style families (--style-family) narrow the primary style pool down to specific directions via a 'category → family' mapping.

rare-style-explorer/scripts/explore_styles.py
"film": {"电影、电视与影像类型"},
STYLE_FAMILIES contains 10 families, with the actual category distribution being: Material 70, Photographic Process 66, Film 66, Animation 64, Graphic 64, Craft 64, Digital 64, Architecture 54, Fashion 54, Toy 54, totaling 620.

Output discipline is documented in SKILL.md: prompts default to Chinese, with English tokens serving solely as metadata; anti-drift negative constraints are mandatory.

rare-style-explorer/SKILL.md
Final prompt variants should be written in Chinese by default. English prompt tokens remain in the style metadata for lookup and manual conversion, but the user-facing prompt text should use Chinese style names, Chinese visual DNA, and Chinese anti-drift constraints.
On the script side, this corresponds to the NEGATIVE constant and the {negative} slot in TEMPLATES.

2Core Capabilities

01Six generation modes: minimal / product / character / poster / scene / material-series, each with its own sentence template
02Targeted exploration across 10 style families (film/fashion/product/photography/illustration/graphic/craft/digital/space/material)
03Reproducible randomness: --seed fixes the random sequence
04Mandatory primary style assignment: --style-id fixes the base style, while other supporting layers continue to vary
05Freshness bias: --freshness high prioritizes newer, less frequent, and more specific entries
06Structured output: --format json directly provides variants + style provenance (including pitfalls / mitigation tips / source batch)
07Manual library query path: SKILL.md provides shell snippets to search the 620 entries by Chinese name, English token, or category
08Fixed chain of anti-drift negative constraints included by default

3External Dependencies

KindDependency / Endpoint
clipython3（3.x，脚本仅用 argparse/json/random/re 标准库）

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Graded/per-use authorization, sandbox recommended
For Administrators
Approval workflow, rollback support, full audit
Evidence & Triggers (File Paths & Excerpts)
  • scripts/: Executes local script or command: python3 scripts/explore_styles.py (bundled script, standard library only; read-only access to bundled style_library.json and prints output)
Audit Notes: Includes a local Python script, but it only performs 'reading the bundled JSON + printing', without external network requests, credential access, local writes, or subprocesses. According to the guidelines, 'only calls local CLI / only local read-write' classifies it as Blue tier. Note: Generation quality and safety rely entirely on the integrity of references/style_library.json (a 10,000-line file that is costly to audit manually; its sha256 checksum can be verified).

5Second-Pass Verification

  • [ok] 620 条宣称是否属实 — json.load 后 len(styles)==620；按类别计数合计 620；description/SKILL.md/README 三处口径一致。
  • [ok] 每个 CLI 选项是否真有实现 — argparse 中逐一存在：mode(choices=sorted(TEMPLATES))、count、seed、style-id、format、freshness、style-family、avoid-generic(BooleanOptionalAction, default=True)；SKILL.md 参数表与实现一致。
  • [ok] 外部资源条目反查 — 唯一依赖是 python3；脚本 import 仅 argparse/json/random/re/pathlib，无第三方包、无 requirements.txt、无网络调用。
  • [ok] 安全结论反例搜索 — 该目录下 explore_styles.py / SKILL.md / README.md / agents/openai.yaml 逐 token 扫描无 os.environ、无 subprocess、无 open(...,'w')；JSON 为纯数据，无可执行片段。
  • [ok] SREF 边界声明是否属实 — 实读 JSON 条目的字段集不含 SREF code；SKILL.md 明示『this skill does not store SREF codes and should not output --sref parameters』，与数据一致。
  • [ok] 去重/加权逻辑是否真的生效（是否有死代码） — pick_unique() 被 make_variant() 调用（base/support/fashion/space 四处），weighted_pick() 被 pick_unique 调用；style_weight 的 freshness/avoid_generic 分支均有调用方传参，非死代码。
  • [ok] 文档与实现的细节出入 — 一处轻微差异：SKILL.md『Combination Rules』给出「一个主风格 + 最多一个表面/光 + 最多一个格式/空间」的散文规则，而脚本 support_for_mode() 在非 material-series 模式下会把多层合并进一次 pick_unique，并在 45% 概率追加 defect 层——即脚本给的辅助层可能多于散文描述，属实现更宽松而非缺失。

6Verdict & Scenarios

  • Reproducible and auditable: The sampling logic is deterministic code; --seed + --format json can reproduce results and trace which style IDs were used for each prompt.
  • Effective anti-generalization: The weighting function reduces entries with high generic word proportions to 0.25 and performs in-batch similarity deduplication to prevent 8 prompt variants from looking identical.
  • Rich data assets (620 entries across 10 categories), each equipped with 'pitfalls / mitigation tips', functioning as a manual curation handbook.
  • Zero external dependencies, runnable offline, no disk writes, with an attack surface close to zero.
    • Best for: Suitable for creators and creator toolchains needing to 'quickly break out of formulaic styles': directional trial-and-error for products/characters/posters/scenes, batch prompt variants, reproducible experimentation (--seed), and cases requiring style provenance in deliverables (--format json); also suitable for automated scheduled generation (README provides an example generating 5 items every 15 minutes).
    Not for: Not suitable for requests requiring fixed oriental mythological fine-brush (Gongbi) styles (which should use shan-ze-school per vsc-distinction boundaries); not suitable for direct final image generation (prompts only); not suitable for workflows requiring Midjourney SREF codes (this skill explicitly does not store or output --sref); not suitable for scenarios with strict style provenance compliance requiring tracing to original gallery authorizations.
    Install Agent-Ready
    ① Verified Mirror 2026-09-15
    Option A · Direct DownloadDownload rare-style-explorer.tar.gz
    sha256: 37bc8ca8da3ceec7…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit fcee81beb6.