Cloud image recognition works by leveraging a server in the cloud that takes on all the heavy-duty processing. First, a user takes a picture with their smartphone and this picture is sent to the server. Second, the server compares that picture with the corresponding database of images. And finally, the server responds to the app with the content (URL, video, etc) set in the content management system (CMS).
On-device image recognition on the other hand can work completely offline. Typically, the app has all the necessary content (e.g. video tutorials) together with the fastest way to search through it: snapping a picture and performing image recognition offline. The On-Device Image Recognition SDK combined with the CraftAR CMS allows you to prepare all the content beforehand and embed it into your app.