Skip to content

Fix rounding when un-premultiplying alpha in getImageData - #2622

Merged
zbjornson merged 2 commits into
Automattic:masterfrom
chiliec:fix/getimagedata-unpremultiply-rounding
Nov 13, 2026
Merged

zbjornson merged 2 commits into
Automattic:masterfrom
chiliec:fix/getimagedata-unpremultiply-rounding

Conversation

@chiliec

@chiliec chiliec commented Sep 20, 2026

Copy link
Copy Markdown
Contributor

Fixes #1933

getImageData un-premultiplies semi-transparent pixels with (int)((float)r * ((float)255 / a)), which truncates. Because cairo's premultiplied channel is already rounded down, truncating again on the way back drops the value by one for most inputs: putImageData([252,252,252,252]) reads back as [251,251,251,252].

This switches to the integer round-half-up division that canvas_unpremultiply_data in src/PNG.h already uses ((c * 255 + a / 2) / a), so getImageData and PNG export un-premultiply identically. Fully transparent/opaque pixels keep the existing fast path.

Exhaustively checking every (value, alpha) pair with alpha in 1..254 (65,024 pixels): exact putImageData → getImageData round-trips go from 1,450 to 17,369, and no pixel gets worse; the remaining differences are the information lost by premultiplication itself.

  • Have you updated CHANGELOG.md?

Validation (linux x64, zig 0.16.0, node 22.12):

zig build
npx mocha test/canvas.test.js -g 1933   # fails before the change: [251,251,251,252]
npm test                                # 306 passing, 6 pending

@zbjornson zbjornson left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Great, thank you.

@zbjornson
zbjornson merged commit 8570a89 into Automattic:master Nov 13, 2026
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Inaccurate RGB values due to a rounding error caused by premultiplication

2 participants