SteerViT: CLS Attention Steering

Choose one of the built-in examples, upload an image, or paste an image URL, then enter a text prompt to try it yourself.

In addition to steering which concepts its global features encode, SteerViT redirects the [CLS] attention toward prompt-relevant regions. DINOv2 is prompt-agnostic and tends to focus on the most salient objects. For more details, check out the project webpage