Techniques

Each technique is a concept unlearning method that modifies a Stable Diffusion model to suppress generation of a target concept. They are distributed as separate installable packages.

Training-based

Technique	Paper	Description
ESD	ICCV 2023	Fine-tunes UNet layers to erase a concept via guided score distillation
MACE	CVPR 2024	Closed-form rank-one edits to cross-attention weights
CA	ICCV 2023	Ablates concept by fine-tuning cross-attention to map it to an anchor
CoGFD	ICLR 2025	Erases concept combinations while preserving individual components
AdvUnlearn	NeurIPS 2024	Adversarially robust unlearning via text-encoder fine-tuning

Inference-time

Technique	Paper	Description
SSD	AAAI 2024	Selectively dampens UNet parameters using diagonal Fisher information
UCE	WACV 2024	Closed-form weight update using concept projection
SAFREE	ICLR 2025	Self-supervised token filtering at inference time
SLD	CVPR 2023	Suppresses concepts via classifier-free guidance manipulation
SAeUron	ICML 2025	Sparse autoencoder feature suppression
Concept Steerers	arXiv 2025	Steers activations away from concept directions at inference
TraSCE	arXiv 2024	Training-free concept erasure via trajectory steering
Free Run	—	Custom Model Evaluation