4 ms·
Can someone explain this better for me and how it would be more useful than a regular transformer?
by brian_herman 5y ago
Can someone explain this better for me and how it would be more useful than a regular transformer?
- robbedpeter 5y agoAttention is more nuanced, operating at different scales. A common flaw in deep fake faces is pupils that aren't quite round, so this type of model could be used to target errors like that. Or if you had a picture of a library bookshelf, the model can attend to the details of individual books, and also attend to the letters of their titles at a per book level.