Repository navigation
[Bug]: test_backend_pgf.py::test_rcupdate fails with texlive 2025 #29790
Description
Bug summary
Hi all, thanks for this useful library! Sadly, with texlive 20250308 test_backend_pgf.py::test_rcupdate image comparison fails in matplotlib 3.10.1. I don't have the images, as the tests ran on an external system.
Code for reproduction
run `test_backend_pgf.py::test_rcupdate`Actual outcome
I am sorry, as I said, I don't have the images and the log is boring otherwise:
[ 503s] ________________________________ test_rcupdate _________________________________
[ 503s]
[ 503s] @needs_pgf_xelatex
[ 503s] @needs_pgf_pdflatex
[ 503s] @mpl.style.context('default')
[ 503s] @pytest.mark.backend('pgf')
[ 503s] def test_rcupdate():
[ 503s] rc_sets = [{'font.family': 'sans-serif',
[ 503s] 'font.size': 30,
[ 503s] 'figure.subplot.left': .2,
[ 503s] 'lines.markersize': 10,
[ 503s] 'pgf.rcfonts': False,
[ 503s] 'pgf.texsystem': 'xelatex'},
[ 503s] {'font.family': 'monospace',
[ 503s] 'font.size': 10,
[ 503s] 'figure.subplot.left': .1,
[ 503s] 'lines.markersize': 20,
[ 503s] 'pgf.rcfonts': False,
[ 503s] 'pgf.texsystem': 'pdflatex',
[ 503s] 'pgf.preamble': ('\\usepackage[utf8x]{inputenc}'
[ 503s] '\\usepackage[T1]{fontenc}'
[ 503s] '\\usepackage{sfmath}')}]
[ 503s] tol = [0, 13.2] if _old_gs_version else [0, 0]
[ 503s] for i, rc_set in enumerate(rc_sets):
[ 503s] with mpl.rc_context(rc_set):
[ 503s] for substring, pkg in [('sfmath', 'sfmath'), ('utf8x', 'ucs')]:
[ 503s] if (substring in mpl.rcParams['pgf.preamble']
[ 503s] and not _has_tex_package(pkg)):
[ 503s] pytest.skip(f'needs {pkg}.sty')
[ 503s] create_figure()
[ 503s] > compare_figure(f'pgf_rcupdate{i + 1}.pdf', tol=tol[i])
[ 503s]
[ 503s] /usr/lib64/python3.11/site-packages/matplotlib/tests/test_backend_pgf.py:142:
[ 503s] _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
[ 503s]
[ 503s] fname = 'pgf_rcupdate2.pdf', savefig_kwargs = {}, tol = 0
[ 503s]
[ 503s] def compare_figure(fname, savefig_kwargs={}, tol=0):
[ 503s] actual = os.path.join(result_dir, fname)
[ 503s] plt.savefig(actual, **savefig_kwargs)
[ 503s]
[ 503s] expected = os.path.join(result_dir, "expected_%s" % fname)
[ 503s] shutil.copyfile(os.path.join(baseline_dir, fname), expected)
[ 503s] err = compare_images(expected, actual, tol=tol)
[ 503s] if err:
[ 503s] > raise ImageComparisonFailure(err)
[ 503s] E matplotlib.testing.exceptions.ImageComparisonFailure: Error: Image files did not match.
[ 503s] E RMS Value: 4.23523981360686
[ 503s] E Expected:
[ 503s] E /home/abuild/rpmbuild/BUILD/python-matplotlib-test-3.10.1-build/matplotlib-3.10.1/result_images/test_backend_pgf/expected_pgf_rcupdate2_pdf.png
[ 503s] E Actual:
[ 503s] E /home/abuild/rpmbuild/BUILD/python-matplotlib-test-3.10.1-build/matplotlib-3.10.1/result_images/test_backend_pgf/pgf_rcupdate2_pdf.png
[ 503s] E Difference:
[ 503s] E /home/abuild/rpmbuild/BUILD/python-matplotlib-test-3.10.1-build/matplotlib-3.10.1/result_images/test_backend_pgf/pgf_rcupdate2_pdf-failed-diff.png
[ 503s] E Tolerance:
[ 503s] E 0
[ 503s]
[ 503s] /usr/lib64/python3.11/site-packages/matplotlib/tests/test_backend_pgf.py:34: ImageComparisonFailure
Expected outcome
tests pass
Additional information
No response
Operating system
openSUSE Tumbleweed
Matplotlib Version
3.10.1
Matplotlib Backend
No response
Python version
No response
Jupyter version
No response
Installation
from source (.tar.gz)
Activity
anupamdwivedi98 commented on Mar 3, 2026
Hi! I would like to investigate this issue.
Since this is an image comparison failure with texlive 2025, I suspect that the PDF rendering output may have slightly changed compared to the existing baseline images.
I will try to reproduce this locally with texlive 2025 and determine whether this requires updating the baseline images or adjusting the PGF backend.
I’ll report back with my findings before opening a PR.
axel-j-larsson commented on Sep 28, 2026
Hi, I also came across this issue when running the tests and would like to look into it.
The test output images appear to show a small difference around the 'h' character in "math":
I ran the test using both texlive 2024 and texlive 2025, and the test only fails when using the latter.
For now, I will try to investigate what is causing the difference, and I'll report back with my findings.
axel-j-larsson commented on Oct 8, 2026
I believe I have found the cause of the failure.
As mentioned, the test passes with texlive 2024 but fails with texlive 2025. I checked whether the difference was coming from Matplotlib itself or from some part of the TeX/PDF rendering pipeline.
I captured the .pgf and .tex files that were generated by backend_pgf.py and compiled the same input manually using both texlive 2024 and 2025. What I found was that the different pdfTeX versions produced small differences in the PDF content. I also checked the Ghostscript rasterization, and that reproduced the same images as the ones generated from the test.
In conclusion, the rendering difference appears to be introduced by changes in pdfTeX between the two texlive versions rather than by the PGF output generated by Matplotlib.
@timhoffm I also tried using the texlive 2025 output as the new baseline image. This makes the test pass for texlive 2025 but then causes it to fail for texlive 2024 instead. As such, I was wondering whether the preferred solution is to update the baseline or leave it as is.
If updating the baseline to the texlive 2025 rendering is preferred, I'd be happy to prepare a PR with the updated image.