ConceptEraser<T>

Models & Types · AiDotNet.Diffusion.Safety

Concept erasure for removing unwanted concepts from diffusion model representations.

For Beginners

Imagine the model’s understanding of concepts as directions in a high-dimensional space. Concept erasure finds the “direction” corresponding to an unwanted concept and removes it, like erasing one axis from a coordinate system. The model can still generate everything else, but it loses the ability to produce the erased concept.

How It Works

Concept erasure removes specific concepts (e.g., artistic styles, identities, or unsafe content) from a diffusion model’s internal representations by projecting embeddings onto the null space of the target concept direction. This prevents the model from generating content related to the erased concept.

Reference: Gandikota et al., “Erasing Concepts from Diffusion Models”, ICCV 2023

Constructors

Constructor Summary
ConceptEraser(double erasureStrength, bool preserveGeneralCapability) Initializes a new concept eraser.

Properties

Property Summary
bool PreserveGeneralCapability { get; } Gets whether general capability preservation is enabled.
double ErasureStrength { get; } Gets the erasure strength.
int ConceptCount { get; } Gets the number of registered concept directions.

Methods

Method Summary
Vector<T> EraseFromEmbedding(Vector<T> embedding) Erases registered concepts from an embedding by projecting onto the null space.
double ComputeConceptPresence(Vector<T> embedding, int conceptIndex) Computes the concept presence score for an embedding (0 = absent, 1 = fully present).
void AddConceptDirection(Vector<T> conceptEmbedding) Registers a concept direction to be erased.