Bug 315522 – [regression] feGaussianBlur in filter is not applied if the stdDeviation contains a 0 in the pair

• Component: SVG
• Status: RESOLVED FIXED

[Nikolas Zimmermann]

Created attachment 479820 [details]
Rendering comparison

LayoutTests/svg/filters/feGaussianBlur.svg used to render fine in WebKit, now MiniBrowser looks different from Firefox/Chrome -- I didn't investigate further.

[Karl Dubost]

Created attachment 479821 [details]
rendering in minibrowser, safari, firefox, chrome

LayoutTests/svg/filters/feGaussianBlur.svg

From left to right
* minibrowser: commit e3c460f943a83d5f57c5adb2326a437a3d98110d (HEAD -> main, origin/main, origin/HEAD)
* STP 244 (WebKit 22625.1.17.19.1)
* Firefox Nightly 153.0a1 (2026-05-20) (aarch64)
* Chrome Canary 150.0.7857.0 (Official Build) canary (arm64)

The most visible effects are on the second and third square.
which if I reduced the SVG corresponds to the 2nd and 3rd in this markup. The first one being 0. 


<svg version="1.1" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink">
<defs>
<filter id="0x0">
    <feGaussianBlur stdDeviation="0"/>
</filter>
<filter id="0x5">
    <feGaussianBlur stdDeviation="0 5"/>
</filter>
<filter id="5x0">
    <feGaussianBlur stdDeviation="5 0"/>
</filter>
</defs>
<g filter="url(#0x0)"  transform="translate(10, 10)">
    <rect x="0" y="0" width="50" height="50" fill="red"/>
    <rect x="25" y="25" width="50" height="50" fill="blue"/>
</g>
<g filter="url(#0x5)"  transform="translate(110, 10)">
    <rect x="0" y="0" width="50" height="50" fill="red"/>
    <rect x="25" y="25" width="50" height="50" fill="blue"/>
</g>
<g filter="url(#5x0)"  transform="translate(210, 10)">
    <rect x="0" y="0" width="50" height="50" fill="red"/>
    <rect x="25" y="25" width="50" height="50" fill="blue"/>
</g>

</svg>


OK found the issue in the markup. Each time stdDeviation attribute contains a `0` (whichever the order), the gaussian blur is not applied.


<svg>
    <defs>
        <filter id="testPASS">
            <feGaussianBlur stdDeviation="0.0001 5" />
        </filter>

        <filter id="testFAIL">
            <feGaussianBlur stdDeviation="0 5" />
        </filter>
    </defs>

    <g filter="url(#testPASS)"
       transform="translate(10, 10)">
        <rect x="0"
              y="0"
              width="50"
              height="50"
              fill="red" />
        <rect x="25"
              y="25"
              width="50"
              height="50"
              fill="blue" />
    </g>
    <g filter="url(#testFAIL)"
       transform="translate(110, 10)">
        <rect x="0"
              y="0"
              width="50"
              height="50"
              fill="red" />
        <rect x="25"
              y="25"
              width="50"
              height="50"
              fill="blue" />
    </g>
</svg>

[Radar WebKit Bug Importer]

<rdar://problem/177906905>

[Karl Dubost]

https://drafts.csswg.org/filter-effects-1/#feGaussianBlurElement

stdDeviation = "<number-optional-number>"
The standard deviation for the blur operation. 

These are the requirements from the spec.

1. If two <number> s are provided, 
  a. the first number represents a standard deviation value along the x-axis of the coordinate system established by attribute primitiveUnits on the filter element. 
  b. The second value represents a standard deviation in Y. 
2. If one number is provided, then that value is used for both X and Y.
3. A negative value or a value of zero disables the effect of the given filter primitive (i.e., the result is the filter input image).
4.  If stdDeviation is 0 in only one of X or Y, then the effect is that the blur is only applied in the direction that has a non-zero value.
5. The initial value for stdDeviation is 0.
6. Animatable: yes.

Are 3 and 4 ambiguous as they seem to contradict each other?

3 says value of 0 disables the effect.
4 says value of 0 on one of the value disables only in the direction of the values. aka stdDeviation="X Y". if X=0 disables along X but Y applies. If Y=0 disables along Y but X applies.

[Karl Dubost]

This was probably regressed by 
https://github.com/WebKit/WebKit/commit/f4ce4a791de7
Bug 315118
https://github.com/WebKit/WebKit/pull/65208

  Root cause — that patch rewrote the early-return guard in FEGaussianBlurSoftwareApplier::apply:

  //  before: AND-zero (both axes must be zero to short-circuit)
  if (!m_effect->stdDeviationX() && !m_effect->stdDeviationY()) return true;

  //  after:  OR-zero (either axis being zero short-circuits)  ← regression
  auto stdDeviation = m_effect->effectiveStdDeviation(filter.renderingOptions());
  if (stdDeviation.isEmpty()) return true;
  
  FloatSize::isEmpty() is width <= 0 || height <= 0 (FloatSize.h:73), so stdDeviation="0 5" and "5 0" now bypass the blur entirely. Software path only —
  Core Image and Skia appliers feed σ straight to their backends and were never affected.

  Fix applied on branch 315522-fegaussianblur-zero-axis (1 file, 1 line):

       auto stdDeviation = m_effect->effectiveStdDeviation(filter.renderingOptions());
  -    if (stdDeviation.isEmpty())
  +    if (!stdDeviation.width() && !stdDeviation.height())
           return true;


A new WPT test should be added for this to catch the regression in the future.

[Karl Dubost]

Pull request: https://github.com/WebKit/WebKit/pull/65642

[EWS]

Committed 313874@main (19ba11bf04d0): <https://commits.webkit.org/313874@main>

Reviewed commits have been landed. Closing PR #65642 and removing active labels.

[Karl Dubost]

Submitted web-platform-tests pull request: https://github.com/web-platform-tests/wpt/pull/60174