Concept-Subspace Score Projection

Estimates a concept's subspace within a diffusion model's score-function representation from contrastive-prompt score differences, then causally edits generation by swapping the projected component of the score function for a new prompt's component, under a provable 'causal separability' condition guaranteeing the swap doesn't leak into other concepts.