wquguru/shisei

---

Content note: the sexy and duo series are suggestive adult pose references (faceless mannequins, no

nudity detail). They stay out of previews and search pages, and the gallery shows them blurred until you tap one

or flip the switch.

Shisei is a self-made pose reference library. Every pose is a faceless mannequin drawn from five angles — front, three-quarter, side, above and below — so you can feed the same pose to an image model from whichever camera you need. Names and categories come in Japanese, Chinese, English and Korean.

All images are original to this project.

What's inside

| Series | Poses shown | Views that passed QC | On the site | |---|---:|---:|---| | Full body · standard (std, P###) | 71 of 73 | 300 / 365 | shown | | Upper body (upper, U###) | 80 of 80 | 359 / 400 | shown | | Full body · sexy (sexy, S###) | 80 of 80 | 352 / 400 | blurred until tapped or switched on | | Duo (duo, D###) | 49 of 55 | 201 / 275 | blurred until tapped or switched on |

Poses rated suggestive (the sexy and duo series, upper's Sexy and Chest emphasis) are blurred in the gallery by default and kept off the marquee, pose pages, share cards and sitemap; tap one to see it, or one switch shows them all.

• Every view is reviewed (docs/qc/). A view that fails (nipple marks, a face or hair,

clothing, background, broken anatomy, the wrong camera angle, a figure that reads as a minor, or a drawing that is not the named pose) is removed from images/ and the catalog, and data/qc.json keeps the reason.

• 43 sets were redrawn by a second generator where the original had at most two usable views or showed another pose

(catalog source: digi, a slightly different mannequin with milder high and low angles; whole sets, never mixed). P038 and P052 still show a different pose from their name and stay hidden.

• Views that were never generated stay empty rather than being filled with a bad drawing.
• 89 photo renders for std P001–P022 and P050 (P019, P020 and P022 are off: they show the old drawing), reviewed like the line art: AI-generated adults, not real people.
• Clay, depth and normal maps for 944 of the line views (270 sets), made from the line art and pixel-aligned with it: a grey mannequin with volume and shading, and depth and normal maps for ControlNet. Reviewed like the rest; the lightbox switches between them.
• data/catalog.json / catalog.csv: series, rating, category and name in four languages, an English pose

description for std, the file for each view, and why a view is empty.

Using the images

Upload a view as the pose or composition reference in your image tool. Using several angles of the same pose keeps the figure consistent across shots. Face, clothing and background come from your prompt or other references.

| Tier | Where | Size | Typical file | |---|---|---|---| | w300 | images/*/w300/ | 300 px wide | ~7 KB | | w600 | images/*/w600/ | 600 × 800 | ~17 KB | | w900 | images/*/w900/ | 900 × 1200 | ~29 KB | | full | original zip | 1344 × 1792 (photos 1120 × 2240, redrawn sets 900 × 1200) | ~100 KB |

The repository holds the three web tiers: about 145 MB for the poses (line art, photos and the clay, depth and normal maps), plus about 150 MB for the costume library. The full-size originals and ready-made zips are on the Releases page; most people only need a zip.

Downloads

v0.4.1, the same files the gallery's download panel offers. Every zip is the 900 px tier of the views that passed QC (the originals excepted), with LICENSE, LICENSE-images and NOTICE; a .sha256 sits next to each. Most people only need the line-art zip. Clay, depth and normal have no originals: they are 900 px to begin with.

| Pack | Contents | Sets | Images | Size | |---|---|---:|---:|---:| | Line art · all series (recommended) | line art + photos | 280 | 1301 | 36.5 MB | | Every format · all series | line art, photos, clay, depth, normal | 280 | 4133 | 69.0 MB | | Clay · all series | clay maps | 270 | 944 | 9.2 MB | | Depth · all series | depth maps, for ControlNet Depth | 270 | 944 | 7.1 MB | | Normal · all series | normal maps, for ControlNet Normal | 270 | 944 | 16.2 MB | | Full body · standard | line art + photos | 71 | 389 | 9.4 MB | | Upper body | line art | 80 | 359 | 10.0 MB | | Full body · sexy (suggestive) | line art | 80 | 352 | 9.3 MB | | Duo (suggestive) | line art | 49 | 201 | 7.7 MB | | Originals | line art + photos at full size | 280 | 1301 | 152.3 MB |

Gallery

web/ is a zero-dependency static gallery built from the catalog: series tabs with categories inside a series, an angle filter, a line-art / photo switch, a by-pose or single-image view, a switch that unblurs the suggestive series, a lightbox with keyboard and swipe navigation and a line-art / photo / compare mode, a page per pose, four UI languages and a dark theme. The spec is in docs/DESIGN.md.

python3 tools/build_site.py --link-images      # writes site/ (images symlinked for a fast preview)
python3 -m http.server 8000 -d site            # http://localhost:8000
python3 -I tools/build_site.py --public --site https://shisei.indaloop.app   # deploy build with images and share cards

Deep links open the lightbox directly: #P001-diag, #P001-diag-photo, #P001-diag-compare; filters live in the query string (?lang=ja&series=std&cat=5&angle=1&src=photo).

Layout

data/       catalog.json, catalog.csv — the source of truth
images/     line/ and photo/, each with w300/, w600/ and w900/ tiers
web/        static gallery source (index.html, app.js, app.css, i18n.js, common.js, changelog.yaml)
tools/      build_images.py, build_site.py, verify.py, import_series.py, qc_merge.py, review_sheets.py, pack_release.py, …
docs/       DESIGN, DECISIONS, GENERATION, LICENSING, translated READMEs, logo, screenshot

Check that the catalog and the images agree:

python3 tools/verify.py

Contributions are welcome: read CONTRIBUTING.md first (coding agents also follow AGENTS.md). Please report security issues privately as described in SECURITY.md.

Status and roadmap

Shisei is at v0.4.0: four series and 280 poses; the catalog is stable and IDs never change. Release notes are in CHANGELOG.md; open questions in docs/DECISIONS.md.

| Milestone | State | |---|---| | 73 standard poses × 5 angles, catalog, web tiers, verify | ✅ v0.1.0 | | Upper body, sexy and duo series, each view reviewed | ✅ v0.2.0 | | Photo renders, after their own QC pass (D-03) | ✅ public | | Public gallery at shisei.indaloop.app (Cloudflare Pages) | ✅ live | | 43 sparse or mislabelled sets redrawn as whole sets (D-09) | ✅ v0.3.0 | | Clay, depth and normal maps for 944 views, aligned with the line art | ✅ v0.4.0 | | P038 and P052: the drawing is another pose | ☐ open | | Fill the P062 below view | ☐ open |

Acknowledgements

The gallery uses Inter Tight, Noto Sans SC and JetBrains Mono (SIL OFL 1.1), served by Google Fonts, and shields.io badges in this README. Licensing rules are in docs/LICENSING.md.

The pose list — which poses, their numbering (orig_id), names and categories — follows the public catalog of ai-pose-gallery by ギア (@giaaibusi); the Chinese, English and Korean names are translations of it. Thank you. Only the list is shared: every image here was made for this project. Shisei is not affiliated with or endorsed by its author.

Notes

Every image was made for this project; none are copied, traced or derived from other pose collections. The figures are mannequins or AI-generated adults and do not depict real people. Do not use them to depict real people or minors, or for explicit content. Image metadata is stripped.

License

• Images (images/ and the Release zips) and the project's own data in data/ are dedicated to the public domain

under CC0 1.0: use, modify and redistribute them, commercially or not, with no attribution needed.

• Code and documentation are licensed under the Apache License 2.0.
• The pose list (selection, numbering, names and categories) comes from ai-pose-gallery's catalog and is credited,

not licensed: neither licence covers it. See NOTICE.

Using the images as a pose reference puts no conditions on the images you generate.