Fix image masks (PDFBOX-6077) - #491
Conversation
A stencil image filled with a pattern draws the paint and the mask into separate scratch images and then combines them. The combine step unconditionally overwrote the paint's alpha with the mask's alpha, so any pixel the pattern itself never painted into (e.g. the gaps between tiles of a tiling pattern) turned opaque black instead of staying transparent. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A soft mask's Paint/PaintContext looks up its backing raster using absolute page-device pixel coordinates, fixed when the soft mask group was rendered. The stencil-mask-with-pattern code renders into an isolated scratch image rather than directly onto the page graphics, so those coordinates no longer lined up and the soft mask silently applied zero alpha everywhere, making the pattern disappear. Unwrap the soft mask, fill the scratch image with its plain underlying paint instead, and apply the soft mask's own alpha afterwards by directly looking up its backing raster through a per-pixel device transform, rather than relying on the Paint/PaintContext machinery that assumed it was rendering onto the real page raster. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
This change produces a white line in the rendering of the file from PDFBOX-5403, in the text "Enzian Immobilien". update: I tried "fix 1" alone and had just that. |
Combining the stencil mask's alpha with the pattern paint's own alpha (instead of overwriting it) exposed a regression on files using a TilingPaint repeated many times over a large area, e.g. one line of text per stencil-masked image: a hairline, fully-transparent seam between adjacent tiles - previously invisible because the mask's alpha always won - now cuts a visible white line through the middle of every line of such text. These seams are a pixel wide at most, caused by sub-pixel rounding at tile boundaries, and are not genuine gaps the pattern never painted into (those remain many pixels wide). Widen the paint's alpha to the maximum of its 4-neighbors before combining it with the mask, which absorbs the hairline seams without meaningfully affecting real gaps. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Root cause: the alpha-combine fix (fix 1) correctly exposes any place where the pattern's own alpha isn't fully opaque — but TilingPaint has hairline (1-pixel), fully-transparent seams at tile boundaries from sub-pixel rounding, which used to be invisible (the old code just overwrote alpha with the mask's, ignoring the paint's). In the PDFBOX-5403 file, a pattern is tiled across many repeated ~14pt-tall bands (one per text line), and the seam lands right in the middle of each line, showing as a light gray/white streak through the text. Fix: before combining, widen the paint's own alpha to the max of its 4-neighbors (a 1px dilate) — absorbs hairline seams while leaving genuinely large gaps (many pixels wide, like the original PDFBOX-6077 case) untouched. |
PDFBOX-6077: Fix stencil masks filled with a pattern
Summary
A PDF image used as a stencil mask (an ImageMask) can be filled with a pattern instead of a solid color. PDFBox renders this case specially: it draws the pattern's paint and the mask into two separate scratch images, then combines them pixel-by-pixel before compositing the result onto the page. Two independent bugs in that combine step caused patterns used this way to render incorrectly.
Bug 1 — mask alpha overwrote the paint's own alpha
The combine step did:
rasterPixel[3] = alphaPixel[0];
unconditionally replacing the paint's alpha with the mask's alpha. Any pixel the pattern itself never painted into — for example the gaps between tiles of a tiling pattern — has alpha 0 in the paint image, but that got overwritten with the mask's (opaque) alpha, turning transparent gaps into solid black.
Fix: combine the two alphas by multiplication instead of overwriting:
rasterPixel[3] = rasterPixel[3] * alphaPixel[0] / 255;
so a pixel is only visible where both the pattern painted something and the mask allows it through.
Bug 2 — soft-masked patterns rendered fully transparent
A pattern can itself have a soft mask applied to it (SoftMask, wrapping the pattern's own Paint). SoftMask looks up its backing (grayscale) raster using absolute page-device pixel coordinates, fixed at the point the soft mask's transparency group was rendered. Because the stencil-mask-with-pattern code fills into an isolated scratch image — not the real page Graphics2D — those coordinates no longer lined up with anything, and the soft mask silently applied zero alpha everywhere, making the pattern disappear entirely.
Fix: unwrap the soft mask, fill the scratch image with its plain underlying paint, then apply the soft mask's own alpha afterward via a new PageDrawer.applySoftMaskAlpha(), which looks up the mask's backing raster directly using a per-pixel device-coordinate transform — correct regardless of the scratch image's resolution — rather than relying on the Paint/PaintContext machinery that assumes it's rendering onto the real page raster. SoftMask gains a few narrow package-private accessors for this.
Commits
Per https://www.apache.org/legal/generative-tooling.html: portions of this PR were
produced with assistance from Claude Code (Anthropic), based on a bug was described in Apache PDFBOX Issue Tracker.
I've reviewed the generated code and confirm to the best of my knowledge that the output does not include any
third-party copyrighted material and is compatible with the Apache License 2.0.