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

fix(tools): put the asset hash in the file name, not a ?v= query

wqugurua514eca01d82
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2b1357ae..b50ffd26 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -23,7 +23,8 @@ the catalog schema.
gallery heading shows on first load.
- A third image tier, `w600` (600 × 800, ~17 KB), feeds the grid on high-density screens, which used to fetch w900:
about 40% fewer bytes per grid image. The repository grows to about 56 MB.
-- Repeat visits load faster: scripts and styles are cached for a year under a content-hash `?v=`, images for a
+- Repeat visits load faster: scripts and styles are cached for a year as `assets/<name>.<content hash>.<ext>` (a
+ hash in the name, not a `?v=` query, which Pages ignores when looking up a file), images for a
week with background revalidation (`site/_headers` for Cloudflare Pages). The hero marquee loads at low priority.
- Images and the project's own data are now CC0 1.0 (`LICENSE-images`); code and docs stay Apache-2.0. Every zip ships
LICENSE, LICENSE-images and NOTICE. README and NOTICE credit the pose list to ai-pose-gallery by ギア (@giaaibusi).
diff --git a/tools/build_site.py b/tools/build_site.py
index f455adb0..5b5de384 100755
--- a/tools/build_site.py
+++ b/tools/build_site.py
@@ -173,12 +173,13 @@ def check_site(out, site):
return bad
-# Cloudflare Pages response headers. Scripts and styles are referenced as ?v=<content hash>, so a cached copy is
-# never stale and can be kept for a year. Image names stay the same when a view is redrawn, so images get a week
-# plus background revalidation rather than immutable. Pages without a rule keep the Pages default (revalidate).
-HEADERS = """/*.js
- Cache-Control: public, max-age=31536000, immutable
-/*.css
+# Cloudflare Pages response headers. Scripts and styles ship as assets/<name>.<content hash>.<ext>: the hash lives in
+# the file name, not a ?v= query, because Pages ignores the query when it looks up a file. With ?v=, an edge still on
+# the previous deployment answered the new URL with the old script, and the zone cache kept that for a year (a blank
+# gallery on 2026-10-29). A hashed name the old deployment lacks cannot be filled with old code, so only assets/ is
+# immutable; the unhashed copies stay for HTML cached before the switch and revalidate like pages. Image names stay
+# the same when a view is redrawn, so images get a week plus background revalidation rather than immutable.
+HEADERS = """/assets/*
Cache-Control: public, max-age=31536000, immutable
/images/*
Cache-Control: public, max-age=604800, stale-while-revalidate=86400
@@ -189,14 +190,19 @@ STAMPED = ("app.css", "app.js", "common.js", "i18n.js")
def stamp_assets(out, ver):
- """Swap the release version in ?v= for each asset's content hash, so an edit between releases busts caches."""
+ """Copy each asset to assets/<stem>.<content hash><ext> and point every page's name?v=<release> reference at it."""
+ (out / "assets").mkdir(exist_ok=True)
+ pages = list(out.rglob("*.html"))
for name in STAMPED:
- digest = hashlib.sha256((out / name).read_bytes()).hexdigest()[:10]
+ data = (out / name).read_bytes()
+ stem, ext = os.path.splitext(name)
+ hashed = f"{stem}.{hashlib.sha256(data).hexdigest()[:10]}{ext}"
+ (out / "assets" / hashed).write_bytes(data)
pat = re.compile(re.escape(name) + r"\?v=" + re.escape(ver) + r"(?=\")")
- for f in out.rglob("*.html"):
+ for f in pages:
doc = f.read_text()
if name in doc:
- f.write_text(pat.sub(f"{name}?v={digest}", doc))
+ f.write_text(pat.sub("assets/" + hashed, doc))
def main() -> None: