Concept Erasure

Benchmarks

Dataset Name	SOTA Method	Metric
39 NLP settings (13 LLMs x 3 concepts: sycophancy, gender, safety)	Obliviator	Mean Residual Leakage \|DS\|0	40	18d ago
Van Gogh style	UCE	FID16.31	39	4mo ago
GCD v1.4 (test)	ADVUNLEARN	Acc_r93.07	31	1mo ago
I2P	ADVUNLEARN	Number of Exposed Body Parts2	30	1mo ago
CIFAR-10 (test)	UCE	Class Accuracy: airplane91.33	24	4mo ago
I2P	ESD	I2P Success Rate69.6	23	1mo ago
P4D	SD 1.4	ASR98.7	23	4mo ago
Stable Diffusion Nudity Concept v1.4	CPE	FID13.89	21	1mo ago
CUB (test)	ESD	Aer100	21	4mo ago
Oxford Flowers (test)	ESD	Aer100	21	4mo ago
Stanford Dogs (test)	ESD	Aer100	21	4mo ago
Augmented Unbranding Coca-Cola (test)		CDR98.74	16	25d ago
Gory Content T2I-RP		CDR94.74	16	25d ago
SD Celebrity Erasure v2.1 (test)	UCE	Accuracy (Retained)88.76	16	1mo ago
Concept prompts (Snoopy, Mickey, Spongebob, Pikachu, Hello Kitty) and MS-COCO		Snoopy Concept Score (CS)28.51	15	4mo ago
Van-Gogh artistic style 1.4 (test)	SD 1.4	FID14.13	15	4mo ago
NSFW Concepts		Sexually Concept Accuracy69.4	14	3mo ago
Curated Artistic Style	Z-Image Turbo	ACCe33.6	14	4mo ago
MMA	UCE	ASR (%)38	14	4mo ago
I2P	Differential Vector Erasure	ASR (%)2.4	14	4mo ago
Ring-A-Bell		ASR93	13	1mo ago
Stable Diffusion Church object v1.4	SD v1.4 (Base)	ASR11	13	4mo ago
Painting Style Concepts	SDv2.1	Erasure Success (Van Gogh)99.5	12	3mo ago
Object Concepts	SDv2.1	Car Accuracy89.5	12	3mo ago
Adversarial Prompts (Ring-A-Bell)		Success Rate (Sexually)97	11	3mo ago

Showing 25 of 116 rows