Summary

Closes #26092

• When passing an array alpha to imshow() with an RGBA image, the alpha values now multiply with the existing alpha channel instead of replacing it
• This makes array alpha behavior consistent with scalar alpha behavior, which already multiplied with the existing alpha channel
• For RGB images, array alpha continues to be used directly as the alpha channel (no change in behavior)

The fix is in _ImageBase._make_image() in lib/matplotlib/image.py. Previously, the array alpha code path unconditionally replaced the alpha channel with A = np.dstack([A[..., :3], alpha]). Now it distinguishes between RGB (use alpha directly) and RGBA (multiply A[..., 3] * alpha).

Before (RGBA + array alpha replaces):

elif np.ndim(alpha) > 0:  # Array alpha
    A = np.dstack([A[..., :3], alpha])

After (RGBA + array alpha multiplies):

elif np.ndim(alpha) > 0:  # Array alpha
    if A.shape[2] == 3:  # RGB: use array alpha directly
        A = np.dstack([A, alpha])
    else:  # RGBA: multiply array alpha with existing alpha
        A = np.dstack([A[..., :3], A[..., 3] * alpha])

Test plan

• Added test_image_array_alpha_rgb - verifies RGB images with array alpha produce same result as manually constructed RGBA
• Added test_image_array_alpha_rgba - verifies RGBA images with array alpha multiply the existing alpha channel
• Added RGB + array alpha case to test_imshow_multi_draw parametrization to verify original array is not modified
• Updated test_interpolation_stage_rgba_respects_alpha_param reference to expect multiplied alpha instead of replaced alpha

🤖 Generated with Claude Code