Rombo.glass implements multiple-scattering GGX BSDF, rough Fresnel, visible distribution of normals and adaptive importance-based sampling for super fast and realistic glasses.
Some image from the comparison video you find below. Renders are not fully converged to better see sampling efforts.
On the left MILA, on the right rombo.glass. You can clearly see how for the same number of iterations sampling is way better for rombo.glass. Render times are faster too (36secs vs 24secs). You can also spot some weird artifacts on the MILA rendering: the horizon is not blurred on the left panel, also it looks like MILA does rough refractions only on first refr level so the back of the sphere looks like it’s not rough but specular, hence the weird overall looking for the left sphere.
Things get worse for MILA when increasing roughness while for rombo.glass thanks to its multiple-scattering the look remains consistent. You can see the MILA weird silhouette and that central ‘cross’ artifact that makes MILA practically unusable with high roughness.
In the following video the full comparison between rombo.glass and MILA.
Simply put. Better visual appearance, faster render times and no weird looking bugs.