Fix BOX downscale dropping boundary source pixels by Dmao233 · Pull Request #9949 · python-pillow/Pillow

Dmao233 · 2026-09-03T01:25:11Z

Changes proposed in this pull request:

BOX downscale assigns a source pixel whose center lands on an output bin edge to exactly one destination pixel
Add regression tests for the 13→6 boundary case and the issue sizes (2755→688, 1837→306)
Note the fix in docs/releasenotes/13.0.0.rst

BOX coefficient ranges used (int)(center ± support + 0.5) plus a discontinuous kernel, so a source pixel whose center landed on an output bin edge could get weight 0 in every output. Tile downscale bins as (left, right] so each source pixel maps to one destination. Co-authored-by: CenFangyu <Dmao233@users.noreply.github.com>

Avoid the deprecated Image.getdata API in the python-pillow#9939 test. Co-authored-by: CenFangyu <Dmao233@users.noreply.github.com>

Co-authored-by: CenFangyu <Dmao233@users.noreply.github.com>

akx · 2026-09-03T13:48:43Z

cc @AlttiRi as the reporter of #9939, you said in that issue you have a BOX down-sampling implementation of your own, so I suspect you have more domain knowledge, can you take a look at this? :)

Co-authored-by: CenFangyu <Dmao233@users.noreply.github.com>

radarhere · 2026-09-13T06:03:00Z

I've created Dmao233#2 with a suggestion.

Took radarhere's Simplify code suggestion.

Dmao233 · 2026-09-14T01:03:55Z

Took the Simplify code suggestion from Dmao233#2.

Keep the BOX downscale note and the HarfBuzz wheel note. Co-authored-by: CenFangyu <Dmao233@users.noreply.github.com>

Dmao233 · 2026-09-22T01:26:35Z

Merged main; releasenotes conflict in 13.0.0.rst resolved (kept BOX note). e7be905.

MohammediYunus · 2026-10-07T16:28:56Z

I checked e7be905 against 5835da6 in isolated native builds. This also fixes a duplication case that would be useful beside the dropped-pixel regressions: shrinking 15 pixels to 14 can count the central source pixel twice on the old implementation.

from PIL import Image

im = Image.new("L", (15, 1))
im.putpixel((7, 0), 255)
out = im.resize((14, 1), Image.Resampling.BOX)
assert [value for value in out.get_flattened_data() if value] == [128]

The base produces two nonzero values, [128, 128]; this head produces [128]. The vertical 1×15 → 1×14 case agrees. That assertion checks exactly one contribution without prescribing which side owns the boundary tie.

On this head, all 59 selftests passed and the focused resampling suite had 103 passes and 2 expected failures. I also verified the small example above on both native builds. I haven’t run the full suite.

Dmao233 · 2026-10-08T01:20:30Z

Thanks, added it as a test for both directions in f1670bf; fails on main, passes here.

MohammediYunus · 2026-10-08T03:01:34Z

Thanks for adding both directions! I rechecked f1670bf: the new regression passes, while the original baseline (5835da6) still duplicates the centre pixel. The resampling suite also passes: 104 passed, 2 expected failures.

radarhere · 2026-10-10T09:19:21Z

What do you think of #10165? It passes all of the tests from this PR, and certainly looks simpler.

Dmao233 · 2026-10-11T00:02:23Z

Checked #10165 against the tests here (boundary drop + 15→14 double-count, both axes): all pass.

Prefer this PR on the approach: BOX downscale should tile (left, right] so each source pixel maps to exactly one output by construction. A 1e-9 nudge on the kernel and rounding is simpler, but it is a tolerance that happens to work for the sizes we hit; it also changes box_filter for every use, not only downscale.

If you still prefer #10165 for the smaller diff, I can close this.

radarhere · 2026-10-22T10:29:09Z

+:data:`PIL.Image.Resampling.BOX` downscaling could skip a source pixel whose
+center landed on an output bin edge, due to range rounding and the discontinuous
+box kernel. Those pixels are now assigned to exactly one output pixel, matching
+the documented "each source pixel contributes to one destination pixel" behavior.


Suggested change

:data:`PIL.Image.Resampling.BOX` downscaling could skip a source pixel whose

center landed on an output bin edge, due to range rounding and the discontinuous

box kernel. Those pixels are now assigned to exactly one output pixel, matching

the documented "each source pixel contributes to one destination pixel" behavior.

:data:`PIL.Image.Resampling.BOX` downscaling could skip a source pixel whose center

landed on an output bin edge, due to range rounding and the discontinuous box kernel.

Those pixels are now assigned to exactly one output pixel, matching the documented

behavior that "Each pixel of the source image contributes to one pixel of the

destination image with identical weights"

I suspect you're quoting https://pillow.readthedocs.io/en/latest/handbook/concepts.html#filters. If so, let's make what's inside the quote marks exact.

Took in 7720f83.

radarhere · 2026-10-22T10:46:17Z

+    /* BOX maps each source pixel to exactly one output with equal weights.
+       When downscaling, (int)(center ± support + 0.5) and the discontinuous
+       box kernel can both miss a pixel whose center lands on a bin edge
+       (issue #9939). Tile bins as (left, right] instead. */


Suggested change

/* BOX maps each source pixel to exactly one output with equal weights.

When downscaling, (int)(center ± support + 0.5) and the discontinuous

box kernel can both miss a pixel whose center lands on a bin edge

(issue #9939). Tile bins as (left, right] instead. */

/* For BOX, each source pixel contributes to one output pixel with equal weights.

When downscaling, (int)(center ± support + 0.5) and the discontinuous box kernel

can both miss a pixel whose center lands on a bin edge (issue #9939).

Tile bins as (left, right] instead. */

Maybe it's just me, but 'maps each source pixel to exactly one output' sounds like a one-to-one mapping, which it isn't.

Took in 7720f83.

Per review feedback on python-pillow#9949.

Dmao233 · 2026-10-23T01:21:07Z

Took both wording suggestions in 7720f83.

Dmao233 and others added 2 commits September 3, 2026 01:23

Use get_flattened_data in BOX boundary regression test

297bd35

Avoid the deprecated Image.getdata API in the python-pillow#9939 test. Co-authored-by: CenFangyu <Dmao233@users.noreply.github.com>

radarhere reviewed Sep 3, 2026

View reviewed changes

Comment thread docs/releasenotes/13.0.0.rst Outdated

radarhere reviewed Sep 3, 2026

View reviewed changes

Comment thread Tests/test_image_resample.py Outdated

radarhere reviewed Sep 3, 2026

View reviewed changes

Comment thread Tests/test_image_resample.py Outdated

Address review: drop issue link, fix lint asserts

b43c670

Co-authored-by: CenFangyu <Dmao233@users.noreply.github.com>

radarhere reviewed Sep 4, 2026

View reviewed changes

Comment thread src/libImaging/Resample.c Outdated

Clarify BOX equal-weight comment in Resample.c

097e00f

Co-authored-by: CenFangyu <Dmao233@users.noreply.github.com>

radarhere mentioned this pull request Sep 4, 2026

Image.BOX resampling skips input pixels #9961

Closed

Simplify code

b7947f7

radarhere mentioned this pull request Sep 12, 2026

Simplify code Dmao233/Pillow#2

Merged

Merge pull request #2 from radarhere/fix-9939-box-resize-edge-pixels

917460a

Took radarhere's Simplify code suggestion.

Dmao233 and others added 2 commits September 18, 2026 01:07

Merge upstream main to resolve python-pillow#9949 releasenotes conflict

c0b1925

Keep the BOX downscale note and the HarfBuzz wheel note. Co-authored-by: CenFangyu <Dmao233@users.noreply.github.com>

Merge upstream main to resolve python-pillow#9949 releasenotes conflict

e7be905

Test BOX downscale does not count a centre pixel twice

f1670bf

radarhere mentioned this pull request Oct 8, 2026

Fix BOX resampling dropping source pixels on window boundaries #10165

Draft

aclark4life added the Bug Any unexpected behavior, until confirmed feature. label Oct 11, 2026

Merge branch 'main' into fix-9939-box-resize-edge-pixels

26e5c03

radarhere requested a review from homm October 22, 2026 10:15

radarhere reviewed Oct 22, 2026

View reviewed changes

docs: match BOX filter quote and soften Resample.c comment wording

7720f83

Per review feedback on python-pillow#9949.

Merge branch 'main' into fix-9939-box-resize-edge-pixels

adedddb

-:data:`PIL.Image.Resampling.BOX` downscaling could skip a source pixel whose
-center landed on an output bin edge, due to range rounding and the discontinuous
-box kernel. Those pixels are now assigned to exactly one output pixel, matching
-the documented "each source pixel contributes to one destination pixel" behavior.
+:data:`PIL.Image.Resampling.BOX` downscaling could skip a source pixel whose center
+landed on an output bin edge, due to range rounding and the discontinuous box kernel.
+Those pixels are now assigned to exactly one output pixel, matching the documented
+behavior that "Each pixel of the source image contributes to one pixel of the
+destination image with identical weights"

Conversation

Dmao233 commented Sep 3, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

akx commented Sep 3, 2026

Uh oh!

Uh oh!

radarhere commented Sep 13, 2026

Uh oh!

Dmao233 commented Sep 14, 2026

Uh oh!

Dmao233 commented Sep 22, 2026

Uh oh!

MohammediYunus commented Oct 7, 2026

Uh oh!

Dmao233 commented Oct 8, 2026

Uh oh!

MohammediYunus commented Oct 8, 2026

Uh oh!

radarhere commented Oct 10, 2026

Uh oh!

Dmao233 commented Oct 11, 2026

Uh oh!

radarhere Oct 22, 2026

Choose a reason for hiding this comment

Uh oh!

Dmao233 Oct 23, 2026

Choose a reason for hiding this comment

Uh oh!

radarhere Oct 22, 2026

Choose a reason for hiding this comment

Uh oh!

Dmao233 Oct 23, 2026

Choose a reason for hiding this comment

Uh oh!

Dmao233 commented Oct 23, 2026

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants