Bug summary

Currently, when upsampling images with interpolation_stage="data", upsampled pixels are set to nan ifany of the underlying data points is nan. This leads to much wider "nan-propagation" than interpolation_stage="rgba".

Code for reproduction

from pylab import *

a = tril(arange(1., 26.).reshape(5, 5))
a[a == 0] = np.nan
axs = figure(layout="constrained").subplots(2, 2)
axs[0, 0].imshow(a, interpolation_stage="data", interpolation="none")
axs[0, 0].set_title("stage=data, interp=none")
axs[0, 1].imshow(a, interpolation_stage="data", interpolation="bilinear")
axs[0, 1].set_title("stage=data, interp=bilinear")
axs[1, 0].imshow(a, interpolation_stage="rgba", interpolation="none")
axs[1, 0].set_title("stage=rgba, interp=none")
axs[1, 1].imshow(a, interpolation_stage="rgba", interpolation="bilinear")
axs[1, 1].set_title("stage=rgba, interp=bilinear")

show()

Actual outcome

Note how the blank area is much wider in the bilinear, data-stage interpolation case.

Expected outcome

Although I'm not sure the choice is objective, I think a blurred boundary (similarly to the bottom right case) would make sense.

Implementation-wise, I suspect this arises from a similar issue as #29711 (comment): it should indeed be possible to interpolate in data space even with nans if we interpret the data array as a single-channel image with an additional alpha channel (0-1, depending on whether the data is nan) and correctly weighting the data by the alpha channel (similarly to the premultiplied alpha filtering suggested in the comment). Without setting a zero weight on the nans, it becomes of course impossible to upsample pixels for which any underlying data points are nan (so setting the upsampled pixel to nan is the only reasonable choice).

Additional information

No response

Operating system

No response

Matplotlib Version

3.11.0.dev525+g9f7b3dd205

Matplotlib Backend

No response

Python version

3.13

Jupyter version

No response

Installation

git checkout