Hi @YogeshCruz - before this is reviewed it may be useful to fix the currently failing tests. Thanks!

What exactly is not working? This works as of 3.10.5

import numpy as np
import matplotlib.pyplot as plt

data = np.array([0, 1]* 50).reshape(10, 10)
alpha = np.linspace(0, 1, 100).reshape(10, 10)

plt.imshow(data, alpha=alpha)

What exactly is not working? This works as of 3.10.5

import numpy as np
import matplotlib.pyplot as plt

data = np.array([0, 1]* 50).reshape(10, 10)
alpha = np.linspace(0, 1, 100).reshape(10, 10)

plt.imshow(data, alpha=alpha)

Thanks @timhoffm for checking it!
The issue I was addressing is about ensuring the shape and dtype of the alpha parameter matches data and gets properly validated especially in edge cases or malformed inputs.

It seems this case works as expected on 3.10.5. I will re-run my tests locally, clean up the failing ones, and update the PR with proper test coverage and code comments to clarify the fix.

Also, I’m facing some local setup issues now previously it worked, but now tests aren’t running due to build errors. I’m working to resolve that and will push a cleaner version soon.

@YogeshCruz please could you post some example code that does not work as expected in v3.10.5? Often when someone thinks they have found a bug, it turns out not to be. It would be a shame to go to the effort of updating your branch if this is one of those cases. If you provide some example code then we can discuss whether there really is a problem.

@YogeshCruz please could you post some example code that does not work as expected in v3.10.5? Often when someone thinks they have found a bug, it turns out not to be. It would be a shame to go to the effort of updating your branch if this is one of those cases. If you provide some example code then we can discuss whether there really is a problem.

You're absolutely right I understand that sometimes what seems like a bug may be expected behavior. Here's a minimal reproducible example that demonstrates the issue I'm seeing in v3.10.5:

import numpy as np
import matplotlib.pyplot as plt

data = np.random.rand(10, 10)
alpha = np.linspace(0, 1, 10).reshape((10, 1))

plt.imshow(data, alpha=alpha)
plt.colorbar()
plt.show()

In this version, passing a 2D alpha array with the same shape as data doesn't apply the per-pixel alpha as expected instead, it raises an error or behaves inconsistently.

If this is intended behavior, I'm happy to close the issue or update the approach accordingly. I just wanted to double-check before proceeding further.

Thanks again for your time and guidance!

Ok, so the issue is that, if the alpha-array size does not match the image size, then funny things can happen.

The fix is to check for the appropriate shape here:

Ok, so the issue is that, if the alpha-array size does not match the image size, then funny things can happen.

The fix is to check for the appropriate shape here:

matplotlib/lib/matplotlib/image.py

Lines 312 to 315 in 40a47e5

Thank you for the clarification!

You're right the core issue arises when the shape of the alpha array doesn't match the image dimensions, leading to unexpected behavior. Adding a shape check in image.py (specifically around lines 312–315) makes sense to catch these mismatches early and provide a clearer error or fallback.

I'll update the branch accordingly to include this validation. Appreciate your guidance on this will follow up once the patch is ready.

I guess it's open to interpretation if that is what people want. The alpha above is one column, so what happens is the first column of the image is shown, and the other 9 are invisible, presumably corresponding to alpha=0 for those cells. I guess I'd prefer an error here, but maybe someone wanted it this way for a reason?

The imshow docs state

If alpha is an array, the alpha blending values are applied pixel by pixel, and alpha must have the same shape as X.

I think I would expect it to either error or broadcast, though the user can of course do their own broadcasting.

I guess it's open to interpretation if that is what people want. The alpha above is one column, so what happens is the first column of the image is shown, and the other 9 are invisible, presumably corresponding to alpha=0 for those cells. I guess I'd prefer an error here, but maybe someone wanted it this way for a reason?

I think raising an error for mismatched alpha array shapes would be safer. The current behavior might work for intentional broadcasting, but it’s more likely to hide mistakes. An explicit check would make the behavior clearer and prevent unintended results.

I think I would expect it to either error or broadcast, though the user can of course do their own broadcasting.

Agreed, either explicit broadcasting or an error would make the behavior predictable. Silent partial application of alpha feels too implicit and could easily mask mistakes.