New Method Unlearns Concepts in Diffusion Models While Retaining Quality

Text-to-image diffusion models face challenges in safely deploying systems due to issues like privacy, copyright, and safety regulations. A critical need exists for "concept unlearning," where specific unwanted concepts can be removed from a trained model without degrading its ability to generate high-quality, diverse, and semantically relevant images for benign prompts. Current unlearning techniques often succeed in forgetting but struggle with retaining overall model performance. A new method, TILDE (TILt-based Distributional Erasure), addresses this by reframing concept unlearning as a distributional alignment task. Instead of simply suppressing the concept, TILDE aims to achieve a minimum-deviation conditional distribution from the original model, ensuring that while the unwanted concept is erased, the relative mass for benign generations is preserved. This is implemented using residual ∇-GFlowNet training to learn the necessary score corrections. Experiments show that TILDE effectively erases concepts across various categories like objects, artistic styles, and characters. Crucially, it significantly improves the retention of general generation capabilities and maintains distributional fidelity, outperforming previous baseline methods.