With Masonry and using the bin packing algorithm/layout, according to your visual requirements, you can (should?) use a system for sizes for the sizing element and get different widths for the underlying columns of the ‘grid’; ie: if the sizing element is a quarter of base width, you scale down some of the widest image to bring more homogeneity-or on the contrary balance it with some enlarged elements that brings some rythm.

Edit: doc has this first example https://masonry.desandro.com/layout that you could use but have to imagine images to be twice the size, similar to a Müller Brockmann grid