Few-step image models raise unsafe output; DICE cuts it
On FLUX Schnell, DICE lowers the unsafe-generation rate to 0.046, versus at least 0.161 for prior erasure methods tested.
Stories per day · 30 days
1 RSS
Relatedautoencoder1calibration1clinical-ml1diffusion-models1distillation1domain-shift1embeddings1fairness1flux1interpretability1language-models1leace1machine-unlearning1minimax1representation-repair1representation-steering1sccr1sdxl1
6 stories, newest first · one card per story
Stories tagged #concept-erasure
Conditional autoencoder erases sensitive concepts and steers them
Under stated assumptions, perfect reconstruction plus independence implies nonlinear guardedness, without task labels.
ConceptACP erases diffusion concepts with sparse autoencoders
On SDXL-Turbo one-step generation, forgetting success rises 36.34 points over the best baseline.
Shared Canonical Component Removal Beyond Expert Views
Researchers introduce SCCR to isolate clinical-model signal that expert feature views do not capture.
Detecting a Shift Is Not Enough: Exact Minimax Limits of Linear Representation Repair
Exact minimax limits show linear shift repair needs far more calibration signal than detection.
HyperErase: Scale-Calibrated Hypernetwork for Multi-Concept Erasure in Text-to-Image Models
HyperErase synthesizes prompt-specific LoRAs to erase multiple concepts from text-to-image models.