why pixel buffers?

[jmccready]

I'd love to understand some of the thinking behind the decision to start implementing ML ops in pixel buffers. I personally like using pixel buffers to do compute operations as well (in my case because i am much more familiar with them and seems to me like its easier to debug when your data can always be viewed as an image) but have never seen anyone else make that same choice. Was getting it to work on more platforms the primary motivation in this case? Are there other advantages you are aware of?