: The variable represents a specific semantic direction that the ACE method attempts to remove or "erase" to prevent the model from generating undesirable images.
In the context of the ACE framework, this "deep feature" likely represents a high-dimensional vector in the model's . Key aspects of these features include: ace.AT_Blacked.1.var
: These features are typically extracted from deep layers of a neural network (such as the last fully connected layer of a pretrained VGGNet or similar architecture) to capture complex abstract information. : The variable represents a specific semantic direction
: ACE introduces learnable gating mechanisms in the model's cross-attention layers, which are fine-tuned per concept using these deep feature representations. : ACE introduces learnable gating mechanisms in the
: The framework uses these features to improve the model's resistance to prompt-based attacks that try to bypass concept erasure.
Deep Feature Consistent Variational Autoencoder - IEEE Xplore