CSS View Transitions Level 2
adds Element.startViewTransition(): a view transition scoped to one element
instead of the whole document. The pseudo-element tree
(::view-transition, ::view-transition-group(), ...) is generated on
the scope element, only its subtree participates, and the rest of the page keeps running --
including other scoped transitions at the same time.
const transition = card.startViewTransition(() => updateCard());
await transition.finished;
90/93 is this build on the current upstream suite
(wpt
rev 536ae94, 77 tests + 16 crashtests), the same revision the WebKit sampler
reports against. The 3 failures: the mathml scope root (needs display: math,
bug 1664261),
occlusion with an opacity: 0 scope (needs IntersectionObserver v2,
bug 1896900), and
display-change-during-transition.html, where the test's
:not(:only-child) qualifier contradicts the spec's pseudo-tree
:only-child matching once auto-nesting re-parents the group (a wpt fix is
in flight; with it the score is 91/93). Fixed since the last capture: svg and replaced
scope roots, negative playback rate, view-transition-name: auto
(bug 1918218), auto-nesting per
csswg-drafts #13420, and the
tentative startViewTransition({ shadowRoots }) option (all 8 tests).
Also fixed: concurrent scoped transitions no longer swap each other's group
keyframes (two transitions sharing a name like root resolved the
first transition's captured geometry, visibly resizing the other scope's
snapshot mid-animation; visible in the previous capture of the clip demo).
80/81 is the wpt snapshot vendored in
firefox-main, the number the in-tree CI runs enforce.
Same build, same demo page, screenshot taken mid-transition with every animation parked at
50% (animation-play-state: paused plus a negative animation-delay).
Before (pref off, the shipping behavior today): the DOM mutation applies instantly, there is
nothing to animate. After: tiles morph inside the scope while the spinner and ticker outside
the scope keep running -- parked at the same angle in both captures, so the only difference
is the transition itself.
The same transition as video: a full cycle (forward transition, settle, reverse transition)
assembled from 62 per-frame captures of the same build with the pref on (animations stepped
through animation-delay, 1.5s duration each way). The before video is the pref-off
behavior: a few frames of the old state, then an instant swap in each direction.
A second capture page exercises the clipping and overflow rules: the pseudo tree escapes
the scope's own bounds but stays clipped by an ancestor overflow: hidden, and
a scope with overflow: auto gains no scroll range from the overlay.




view-transition-name comes from an adopted stylesheet. The dynamic
::view-transition styles resolve in the shadow root's style scope, so the part morphs from
square to circle across the card.

overflow: hidden still cuts it at the edge. Mid-flight the circle is visibly
outside the scope and clipped by the ancestor.

overflow: auto with styled (non-overlay)
scrollbars. The flying group leaves the content box during the morph, yet no scrollbars
appear and scrollLeft/scrollTop stay 0,0: the overlay is excluded from the
scope's scrollable overflow.
One feature-flagged implementation pass, developed as a 12-commit branch on top of
firefox-main and being split into Phabricator revisions under
bug 2075158:
the Element.startViewTransition() API and scoped lifecycle, the pseudo-element
tree on arbitrary scope elements, the snapshot containing block and its layout (no
scrollable-overflow contribution to the scope), clip escape past the scope with ancestor
clips still applied, scoped capture and paint order, dynamic UA styles resolved in the
scope's (possibly shadow) style scope, concurrent scoped transitions, hit testing over the
overlay, devtools animation metadata, and UA-sheet parsing of the new pseudos with the pref
off.
| before | after | |
|---|---|---|
| wpt css/css-view-transitions/scoped (81 tests, vendored snapshot) | suite skipped, 0 passing | 80 of 81 pass, 63 of 66 subtests; 1 known failure needs view-transition-name: auto |
| full css-view-transitions suite | as expected | as expected, 0 unexpected results, no document-transition regressions |
| wpt scoped suite at upstream rev 536ae94 (93 tests) | suite skipped, 0 passing | 90 of 93, including the 8 tentative shadowRoots-option tests; remaining failures are two platform gaps (display: math, IntersectionObserver v2) and an upstream test bug |
| WebKit implementation (PR #81741, same suite rev) | 87 of 93 at the same wpt revision: WebKit implements the tentative shadowRoots option, Gecko passes scoped cases WebKit still misses. The WebKit sampler runs the same four cases with the same capture recipe: two independent implementations, same output. | |
# test counts (CI reftests are not headless, so neither is this)
xvfb-run -a ./mach wpt testing/web-platform/tests/css/css-view-transitions/scoped --log-raw scoped.log
xvfb-run -a ./mach wpt testing/web-platform/tests/css/css-view-transitions --log-raw all.log
# 90/93: overlay the 16 files added upstream since the vendored snapshot
# (github.com/web-platform-tests/wpt rev 536ae94) into the working tree,
# re-run the directory with --manifest-update, score wpt.fyi-style
# (test passes if PASS, or OK with zero failing subtests), restore the tree
# captures: same build, marionette drives a capture page that parks every
# animation at frame t/24 via animation-play-state: paused and a negative
# animation-delay; the pref is the only variable between before and after
xvfb-run -a ./mach python capture.py # obj-debug/dist/bin/firefox, 800x600, ffmpeg x11grab
# videos from the 25 per-frame captures
ffmpeg -framerate 16 -i after-%02d.png -c:v libx264 -pix_fmt yuv420p after-main.mp4
ffmpeg -framerate 16 -i after-%02d.png -vf format=yuv420p -c:v libvpx-vp9 -b:v 0 -crf 32 after-main.webm