Amil Dravid
@_amildravid
PhD @Berkeley_AI
ID: 1639674617378795520
http://avdravid.github.io 25-03-2023 17:04:54
170 Tweet
601 Followers
459 Following
Artifacts in your attention maps? Forgot to train with registers? Use π©ππ¨π©-π©ππ’π π§ππππ¨π©ππ§π¨! We find a sparse set of activations set artifact positions. We can shift them anywhere ("Shifted") β even outside the image into an untrained token. Clean maps, no retrain.
