Skip to content

💃 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

Merge branch 'feat/v3-fill' into feat/costume-tray

wqugurua3e8566d028b
diff --cc AGENTS.md
index 8260024b,74064f45..2facb27f
--- a/AGENTS.md
+++ b/AGENTS.md
@@@ -67,13 -66,12 +68,15 @@@ images/line/w300/ grid tier for 1x s
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 and alpha/w900/ transparent (D-10),
costume_<ID>_<view>.webp; QC in data/costume-qc.json from data/costume-qc-input/
+ images/{prop,furniture}/ grey/{w300,w600,w900}/ on #D9D9D9 and alpha/w900/ transparent, 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)
+ 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 and changelog
per language (zh at the root, ja/ en/ ko/), og/ share cards, sitemap, robots, 404, manifest,
_headers (Cloudflare cache rules)
diff --cc CHANGELOG.md
index b513870e,7501a202..18942adf
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@@ -8,94 -8,7 +8,95 @@@ the catalog schema
## [Unreleased]
### Added
+- Costumes on the site: a costume library page per language (`costume/`, 484 items with a passing view, grey tiers
+ only; the transparent versions stay in the download packs) with category, gender and view filters, a four-view
+ lightbox and lingerie blurred like the suggestive poses. A pairing tray (`web/tray.js`) keeps one pose and one
+ costume across both pages, maps the pose angle to the costume view and gives the combined prompt and both images
+ (phones: one share sheet). The pose lightbox gets "Pick a costume for this pose"; the home page a band after the grid.
+
+### Changed
+- Performance: fonts are self-hosted Latin subsets (Inter Tight, JetBrains Mono; 65 KB, OFL files in `web/fonts/`)
+ and CJK text uses system fonts, replacing Google Fonts (28 requests, 1.6 MB, often unreachable from mainland China).
+ The lightbox fetches what is on screen first and preloads only afterwards, two at a time, cancelled on every step;
+ overview tiles pick 300 or 600 px by screen. Re-renders keep existing nodes and swap decoded images, so changing angle
+ no longer flashes; failed images offer a retry. Marquee tiles past the first eight load lazily (25 images on first
+ load instead of 53); grid cards use content-visibility.
+- Phones: the overview's button is "Save this angle" and hands every format of the angle (up to five images) to
+ the share sheet at once ("Save 5 Images" on iOS); an enlarged format saves just that image ("Save clay"). Browsers
+ that cannot share files, such as WeChat, show "How to save" with a long-press hint. Desktop is unchanged.
+- The lightbox preloads what the next step needs once the shown images are in: the overview's other angles, an
+ enlarged format's other angles and this angle's other formats, and the next pose's opening view. Low priority,
+ skipped on Save-Data and 2G connections.
+- Enlarged formats get a format row under the image: "All" (the overview) and each format of the set, one tap to switch;
+ it replaces the "← All formats" buttons. On phones, swiping up and down on the image steps through the same row and
+ flashes the format name. The phone overview shows all five images (three columns, the photo filling the third), and
+ the "This pose also has a photo" row is gone. Marquee tiles still outline their format, without the label.
+
+### Fixed
+- Costume library: the lingerie category's repair round brings it to 1889 of 1976 views and 484 of 494 items
+ (was 1870 and 482). Background left inside cut-outs and leg openings by the matte is cleared, and pale floating
+ pieces it dropped (white cuffs, a collar) are restored, with the new `tools/costume_holes.py`.
+- Phone lightbox cramped on first open: the box is 100dvh, about 180 px shorter while the address bar shows, and only the
+ image gave way. The meta line, angle row and format note are trimmed on phones so the image keeps its height.
+
+- The lightbox opens on an overview of the angle: line art, clay, depth and normal in a 2×2 grid and the photo in a
+ column beside it (a row under it on phones). Click a tile to see that format alone, "← All formats" goes back. The
+ format bar is gone; sets with only line art open on the line art. `#P001-diag` is the overview, `-line` and the other
+ format suffixes open one format, and the old `-compare` links land on the overview. Strings in all four languages.
+
+### Changed
+- Back to the overview is a black "← All formats · Clay" pill on the stage (desktop) or the first button of the bottom
+ bar (phones); Esc and the browser's back go to the overview before closing. Marquee tiles open the overview with
+ their format outlined and a short-lived "The one you clicked" label. Strings in all four languages.
+
+### Fixed
+- Lightbox ← → sometimes did nothing: opening a map format focused the line-art slider first, which took the keys.
+ The close button now takes the first focus and the slider no longer swallows the arrow keys.
+- Angle thumbnails for a view without clay, depth or normal said "undefined"; they show a layers icon like the line
+ and photo gaps. Switching format on such an angle stays there and shows the gap box ("斜侧 · 暂无白模") with the
+ ways out (this angle's line art, the nearest angle with that format) instead of jumping to another angle.
+
+### Added
+- Home marquee: a tile opens its pose in the lightbox at that angle and in that format (a depth tile opens depth).
+ The rows pause on hover as before and stay out of the tab order; the grid remains the keyboard way in.
+
+### Changed
+- Chinese hero title 姿势参考 · 五机位: the site has five formats now, so the title names none of them; 线稿 stays in
+ the line below and in the page title.
+- README: the tagline in all four languages mentions the clay, depth and normal maps; the repository size note covers
+ the maps and the costume library (it said 63 MB); a new screenshot of the hero with the mixed-format marquee.
+ AGENTS.md describes the maps in §1.
+- Page titles, meta descriptions and home share cards in ja / en / ko (and the zh description) mention the clay,
+ depth and normal maps, so search results show them; the home card subtitle breaks only between phrases.
+ AGENTS.md §6 gains the format names per language and the wording rules (白模 is only the clay format).
+
+## [0.4.1] - 2026-10-31
+
+### Changed
+- Downloads say what each zip holds. The panel leads with "Line art · all series" (line art and photos, chips with
+ the counts), then a new `shisei-complete` zip with every format (69 MB), one zip per map format
+ (`shisei-clay`, `shisei-depth`, `shisei-normal`: depth and normal were one `shisei-controlnet` zip, but people pick
+ one ControlNet and the normals are twice the size), the series zips marked line art + photos, and the originals
+ marked line art + photos at full size, with a note that the maps have no originals. The hero download button no
+ longer shows the line-art zip's size as if it were everything. README download tables gain a contents column.
+- Chinese copy: 白模 now only names the clay format; the mannequin line art is 素体 (hero title 姿势线稿参考, page
+ title, share card, pose-page descriptions, README). Every language's hero line mentions the three map formats.
+
+## [0.4.0] - 2026-10-31
+
+### Added
+- Clay, depth and normal maps for 944 line views (270 sets), made from the line art and pixel-aligned with it
+ (Qwen-Image-Edit clay, Depth Anything V2 depth, Marigold normals; docs/GENERATION.md). Catalog `maps: {view: file}`,
+ `images/{clay,depth,normal}/` in the three tiers, QC round 5 in qc.json `maps`: an alignment score
+ (`tools/map_align.py`), a five-panel review (`review_sheets.py --mode maps`, `docs/qc/maps_prompt.md`) and a clay
+ close-up for nipple marks (`--mode mapclay`, `maps_nipple_prompt.md`); `build_images.py maps=…`, `verify.py` and
+ `qc_merge.py` handle them. Of 1,212 reviewed views, 268 fail (misaligned upper-body renders, modelled faces or nipples).
+- Lightbox: the line / photo / compare switch adds clay, depth and normal, each with a one-line note; a slider wipes the
+ line art in over a map. Downloads get `_clay` / `_depth` / `_normal` names, deep links `#P001-diag-clay`. The home
+ marquee mixes the three formats in, labelled. The grid and its filters are unchanged. Strings in all four languages.
+- Release zips `shisei-clay` and `shisei-controlnet` (depth + normal), all series, w900; the download panel and the
+ READMEs list them under "Other formats".
+ - Scene, prop, furniture, light and animal libraries, first run (D-12–D-17; the gallery pages come later): 1328 views generated by Muse from `series_brief.py`, taken in with `series_intake.py` (props and furniture matted onto `#D9D9D9` like costumes), reviewed per view by Sonnet with `docs/qc/{scene,prop,furniture,light,animal}_prompt.md` and merged by `series_qc.py`. 895 views pass: scenes 168/204, props 114/144, furniture 163/240, lights 108/165, animals 342/575; WebP tiers in `images/<lib>/` built by `series_build.py`, checked by `verify.py`.
- Costume library data and images (D-10, D-11; the gallery pages come later): 494 costume cards in 12 categories in
`data/costumes.json`, generated by Muse as ghost-mannequin product photos in four views, matted locally and QC'd
per view. 1870 of 1976 views pass (482 items); WebP on the house grey `#D9D9D9` in three tiers plus a transparent