💃 Free CC0 mannequin pose reference library for AI image tools (Nano Banana, ChatGPT, ComfyUI, Midjourney) — 250+ poses × 5 angles in line art, names in ja/zh/en/ko
AGENTS.md — Shisei
The single guide for AI agents (Claude Code, Codex, …) and humans working in this repository. CLAUDE.md only
includes this file.
1. What this is
A self-made pose reference library: faceless mannequin line art, each pose drawn from five angles (front, diag,
side, above, below), for people who drive AI image tools with pose references. Four series: std full body
(P###, 73 sets), upper upper body (U###, 80), sexy full body (S###, 80) and duo two figures (D###, 55); the last
two are rated suggestive. Plus photo renders for 23 std sets, clay / depth / normal maps for most passed views (made
from the line art and aligned with it), a four-language catalog (ja / zh / en / ko) and a static gallery in web/.
2. Hard rules
- No outside images. Never download, cache, mirror, re-encode or trace images from ai-pose-gallery (ギア @giaaibusi), its mirrors or any other pose site. Its terms forbid redistributing the image data, processed copies included; a mirror's open-source licence covers its code at most. Do not name or link the mirrors in the repo or on the site. UI patterns may be referenced; code and copy may not be copied. The pose list follows its catalog and is credited in README and NOTICE, not licensed (D-01).
- Catalog is the source of truth. Every image is listed in
data/catalog.json;tools/verify.pymust pass. A view that could not be generated isnullwithgaps: missing— never fill a gap with a mirrored, flipped or low-quality drawing. - No view ships without a QC pass. Every catalog view has
status: passindata/qc.json; a new or redrawn view goes through docs/qc/README.md (review sheets → Sonnet review + pose-name check →qc_merge.py) before it entersimages/; photo renders likewise need apassin qc.jsonphotos(docs/qc/photo_prompt.md). Never mark a viewpassby hand to get it in. - Ratings. Every pose has
rating(general/suggestive). Everything that passed QC is public; suggestive poses are blurred in the gallery by default (tap one to see it, one switch shows all) and stay out of the marquee, pose pages, share cards and sitemap (D-06). Nipple marks, exposure, genital detail and a figure that reads as a minor are generation errors: they fail QC. A per-pose rating change is a human decision recorded indata/rating_overrides.json, not a catalog edit:import_series.pyapplies it on re-import andverify.pychecks it. Costumes follow the same rules; the LG (性感内衣) category is allsuggestiveand its boundary is D-11: no sheer fabric, no cut-outs or openings at the nipple or crotch, no restraints, nothing school-coded or minor-coded, no police or military, no real insignia. - IDs are stable.
P001… never get renumbered or reused (P023 and P028 are retired). New poses take the next free number; the source-list number goes inorig_id. - Full-size originals stay out of git. Commit only the
w300,w600andw900tiers; downloads ship as Release zips (tools/pack_release.py: w900 per series + all, recommended; originals), sizes indata/packs.json. - Content safety. Figures are mannequins or AI-generated adults. Nothing that reads as a minor, a real person or explicit content enters the repo.
- Do not publish the repo, create releases, deploy, post or spend money without a human's OK (§8).
- Never weaken
verify.pyto make it pass; never deletedocs/content to "simplify" — fix it when outdated.
3. Layout
AGENTS.md this file; CLAUDE.md only includes it
README.md overview (English); docs/readme/README.{zh-CN,ja,ko}.md — keep all four in step
LICENSE, NOTICE Apache-2.0 for code and docs; LICENSE-images: CC0 for images and data (D-01); all three in every zip
CONTRIBUTING.md contributor guide; SECURITY.md, CODE_OF_CONDUCT.md; .github/ issue + PR templates
CHANGELOG.md Keep a Changelog; every user-visible change adds a line under [Unreleased]
data/catalog.json 288 poses: id, orig_id, series, rating, added_in, class_*, name_* (jp/cn/en/kr), desc_en,
views{front..below} (file or null), gaps{view: missing|line|structure}, photos{view: file},
maps{view: file} (clay / depth / normal of that line view),
bodies (duo only: MF / FF), source (redrawn sets only: digi, D-09),
fills{view: source} (single views redrawn later by the original generator)
data/packs.json Release download zips: file, bytes, sets (tools/pack_release.py; the gallery shows the sizes)
data/qc.json per-view and per-photo QC status, reason and file name; per-set pose-name notes; sources of
redrawn sets (docs/qc/README.md)
data/qc-input/ every reviewer result behind qc.json; `qc_merge.py data/qc-input` rebuilds it exactly
data/rating_overrides.json per-pose rating decisions that win over the category rule (human-decided)
data/costume-src/ costume library source (D-10): _categories.json + one <CAT>.json of forms × colorways
data/costumes.json costume items C###, expanded by tools/costume_cards.py; views front/back/side/diag
data/{scene,prop,furniture,light,animal}-src/ scene / prop / furniture / light / animal libraries (D-12–D-17, go-ahead
2026-11-01): category rules + cards, expanded by tools/series_cards.py into
data/{scenes,props,furniture,lights,animals}.json (E### / T### / F### / L### / A###);
QC in data/<lib>-qc.json from data/<lib>-qc-input/ (tools/series_qc.py)
data/catalog.csv tools/export_csv.py output (o = passed, x = failed QC, - = not generated)
images/line/w300/ grid tier for 1x screens and thumbnails, 300 px wide WebP q70
images/line/w600/ grid tier for 2x / 3x screens, 600×800 WebP q72
images/line/w900/ lightbox/download tier, 900×1200 WebP q75
images/photo/… same two tiers for photo renders
images/{clay,depth,normal}/ maps of passed line views (catalog `maps`), same three tiers and file names as the line
art, made from it and pixel-aligned (docs/GENERATION.md); QC in qc.json `maps`
images/costume/ costume views: grey/{w300,w600,w900}/ on #D9D9D9 (D-10; no transparent tier, D-18),
costume_<ID>_<view>.webp; QC in data/costume-qc.json from data/costume-qc-input/
images/{prop,furniture}/ grey/{w300,w600,w900}/ on #D9D9D9, as costumes
images/{scene,light,animal}/ plain/{w300,w600,w900}/ as delivered (pose tier quality); <lib>_<ID>_<view>.webp
web/ gallery source: index.html template, app.js, app.css, i18n.js, common.js (share sheet, copy,
language links), favicon.svg, changelog.yaml (reader-facing changelog, four languages), fonts/ (self-hosted
Latin subsets of Inter Tight and JetBrains Mono with their OFL files; CJK uses system fonts)
site/ build output of tools/build_site.py (git-ignored; deploy this): home, pose pages, library pages,
library item pages (<lib>/<ID>/, general only) and changelog
per language (zh at the root, ja/ en/ ko/), og/ share cards, sitemap, robots, 404, manifest,
_headers (Cloudflare cache rules)
build/og-cache/ share-card render cache (git-ignored)
tools/ build_images.py, build_site.py (+ site_pages.py, og_cards.py), verify.py, pack_release.py,
import_series.py, import_redraws.py, import_fills.py, import_descs.py, review_sheets.py, line_dups.py, qc_merge.py, rating_check.py,
export_csv.py, costume_cards.py, costume_brief.py (docs/costume/PLAN.md),
series_kit.py + series_cards.py + series_brief.py, series_intake.py + series_sheets.py +
series_qc.py + series_build.py (scene / prop / furniture / light / animal libraries)
docs/qc/ QC method, results and the reviewer prompts (panels, name, limbs, proportion, photo, rating,
costume, scene, prop, furniture, light, animal)
.claude/agents/ qc-reviewer: the Sonnet sub-agent every QC batch is sent to
docs/{scene,prop,furniture,light,animal}/PLAN.md plans and run results for the five new libraries (D-12–D-17)
docs/DESIGN.md gallery layout, tokens, interaction spec (agent-facing copy of the design canvas)
docs/DECISIONS.md open questions and decisions awaiting a human
docs/GENERATION.md how the sets were made, gap log
docs/LICENSING.md licensing rules (outside images, redistribution, AI-generated images)
docs/logo.svg, docs/screenshot.png README assets; retake the screenshot when the hero changes
.github/workflows/ CI: verify.py
File names are <series>_<ID>_<view>.webp (std_P001_front.webp, upper_U001_front.webp), identical across tiers.
All series share data/catalog.json; a new series gets its own prefix and ID letter.
4. Environment
- macOS, zsh. Unquoted
$varis not word-split in zsh; write flags literally or use arrays. cwebp1.6 (Homebrew) for image tiers; Python 3.9+ (system) is enough for the tools; run Python with-I.- The source set lives outside the repo (e.g.
~/Downloads/pose-gallery-std73/); treat it as read-only input.
5. Common commands
python3 tools/build_images.py std=~/Downloads/pose-gallery-std73 sexy=~/Downloads/pose-gallery-sexy80 \
upper=~/Downloads/pose-gallery-upper80 duo=~/Downloads/pose-gallery-duo55 \
digi=~/Downloads/pose-gallery-digi66 fill-r1=~/Downloads/pose-gallery-fill-r1 \
maps=~/Downloads/pose-gallery-clay # passing views only; prunes the rest
# digi = redrawn sets (D-09); fill-r1 = single views refilled by Muse; maps = clay / depth / normal
python3 tools/verify.py # catalog ↔ qc.json ↔ images, gaps, orphans
python3 tools/import_series.py SRC --series X --added 0.2.0 [--general Cute] # merge a new series
python3 -I tools/import_redraws.py SRC OUT --source digi --only P019,… # whole-set redraws, then QC (D-09)
python3 -I tools/import_fills.py ~/Downloads/muse-fill-r1-batchNN ~/Downloads/pose-gallery-fill-r1 --source fill-r1
# single-view fills from Muse, then QC (docs/GENERATION.md)
python3 -I tools/import_descs.py ~/Downloads/pose-prompts-export # desc_en from the generator's own prompts
python3 tools/qc_merge.py data/qc-input # apply review results (docs/qc/README.md)
python3 tools/build_site.py --link-images && python3 -m http.server 8000 -d site # preview the gallery
python3 -I tools/build_site.py --public --site https://shisei.indaloop.app --media https://shisei-media.pages.dev/
# deploy build (D-05): site/ to Pages project shisei, site-media/ (library images) to shisei-media; --no-og skips cards
python3 -I tools/pack_release.py 0.2.0 std=… upper=… sexy=… duo=… digi=… # dist/*.zip + sha256, data/packs.json
6. Way of working
- Adding poses or views: drop sources into the source set →
import_series.py(or edit the catalog: all four languages) →import_descs.pyfordesc_en, the generation prompt (verify requires one) → QC per docs/qc/README.md (panels + name; limbs, proportion and rating on what passed) →qc_merge.py data/qc-input→build_images.py→verify.py→export_csv.py→ CHANGELOG line and a GENERATION.md note. - Translations: ja is the original naming; zh / en / ko are translations. Keep names short (
動作・手の位置shape); change all four together. - Formats and wording (2026-10-31): the site has five formats: line art, photos, and clay / depth / normal maps. In Chinese, 白模 names only the clay format; the faceless mannequin is 素体 (never 白模(素体)). Format names: zh 线稿 / 实拍 / 白模 / 深度 / 法线, ja 線画 / 実写 / クレイ / 深度 / 法線, en line art / photo / clay / depth / normal, ko 선화 / 실사 / 클레이 / 깊이 / 노멀. Hero titles name no single format (zh 姿势参考); the hero line, page title, meta description, share card and README taglines say what the library holds in all four languages, so a new format or count changes all of them together. Describe the maps by what they are (clay: usable as a reference image; depth and normal: for ControlNet) and never claim an effect nobody has tested. Every download names its contents.
- Web: follow docs/DESIGN.md. Every UI string lives in
web/i18n.jsin all four languages;index.htmlplaceholders (__CATALOG__,__REPO__,__VERSION__) are filled bybuild_site.py. Zero runtime dependencies, works at 390×844 without horizontal scroll, every string in all four languages, images lazy-loaded withalt= ID + name + view, keyboard: ← → angle, ↑ ↓ pose, Esc closes. - Pages and SEO: strings the build needs (page titles, share sheet, changelog) live in the strict-JSON
I18N_PAGESblock at the end ofweb/i18n.js. Reader-facing updates go inweb/changelog.yaml(all four languages; the build audits them) as well as CHANGELOG.md. Share what has a page: pose pages, not#P001hashes. Only std / general-rated poses with a usable view get pose pages, cards and sitemap rows. - Comments and docs: code comments in English, explain why.
docs/may be Chinese; README is English-first with translated copies.
7. Git
- Conventional Commits in English:
feat(data): add 6 squatting poses,fix(images): rebuild w900 for P034,docs: …. Scopes:data,images,web,tools,docs,ci. - One logical change per commit; image rebuilds go in their own commit so diffs stay reviewable.
- Default branch
main. Tag releasesvX.Y.Zand attach the full-size zip.
8. Autonomy limits
Agents may edit files, rebuild tiers, run tools and commit locally. A human must OK: pushing to a public remote,
creating a release, deploying web/, choosing or changing a licence (docs/DECISIONS.md D-01), deleting images, or
adding a new image series.
9. Roadmap
| Item | State |
|---|---|
| 73 standard poses, catalog, two web tiers, verify | ✅ v0.1.0 |
| Gallery design (desktop, lightbox, mobile, card states, logo directions) | ✅ design canvas |
web/ static gallery generated from the catalog |
✅ |
Pages deploy of site/ to shisei.indaloop.app (D-05) |
✅ OK'd 2026-10-29 |
| ja / ko README copies, screenshot, community files | ✅ |
| Licence (D-01): code Apache-2.0, images and data CC0 | ✅ 2026-10-29 |
| Photo policy (D-03): public after photo QC | ✅ 2026-10-29 |
| Fill P062 below | ☐ open |
| upper + sexy series, line-art QC, series / rating filters | ✅ v0.2.0 (2026-10-29) |
| duo series (55 sets) and photo QC | ✅ v0.2.0 (2026-10-29) |
| Relabel the 23 std sets whose drawing is another pose (D-07) | ◐ 21 redrawn (D-09); P038, P052 open |
| Redraw gaps from a second source, whole sets only (D-09) | ✅ 43 sets, 2026-10-29 |
| Rewrite failed views out of git history before going public (D-08) | ✅ 2026-10-29 |
| Clay, depth and normal maps for passed line views, QC round 5 | ✅ v0.4.0 (2026-10-31) |