Transformers also make for a damn good base to graft just about any other architecture onto.
Like, vision transformers? They seem to work best when they still have a CNN backbone, but the "transformer" component is very good at focusing on relevant information, and doing different things depending on what you want to be done with those images.
And if you bolt that hybrid vision transformer to an even larger language-oriented transformer? That also imbues it with basic problem-solving, world knowledge and commonsense reasoning capabilities - which, in things like advanced OCR systems, are very welcome.
Like, vision transformers? They seem to work best when they still have a CNN backbone, but the "transformer" component is very good at focusing on relevant information, and doing different things depending on what you want to be done with those images.
And if you bolt that hybrid vision transformer to an even larger language-oriented transformer? That also imbues it with basic problem-solving, world knowledge and commonsense reasoning capabilities - which, in things like advanced OCR systems, are very welcome.