Bindings are good until they are not. Typically for my own tensor library I used bindings for optimized matrix multiplication but people have difficulties installing those dependencies on Linux.
Also for Nim, you could do a lot of precomputation at compile-time which you cannot do for bindings. Many of the recent languages: Go and Java but even Python have their own pure image manipulation library. And lastly, if you want to do image manipulation in production, sure use OpenCV, but a lot of Nim developers are here because they find the language fun and they don't have something to ship in 2 months but they have time and the willingness to learn both a language and an application domain. It's a lot more fun (usually) and obviously better for learning to write something from scratch than to wrap something already ready.
