Subtracting the text embedding of an unwanted concept from a target word's embedding, with cross-attention keys and values steered in opposite directions, suppresses strongly entangled content in Stable Diffusion and personalized models.
Erasing concepts from diffusion models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Translation of Text Embedding via Delta Vector to Suppress Strongly Entangled Content in Text-to-Image Diffusion Models
Subtracting the text embedding of an unwanted concept from a target word's embedding, with cross-attention keys and values steered in opposite directions, suppresses strongly entangled content in Stable Diffusion and personalized models.