notesum.ai

Published at November 15

Safe Text-to-Image Generation: Simply Sanitize the Prompt Embedding

cs.CR

cs.AI

cs.CL

Released Date: November 15, 2024

Authors: Huming Qiu¹, Guanxu Chen¹, Mi Zhang¹, Min Yang¹

Aff.: ¹Fudan University, China

Arxiv: http://arxiv.org/abs/2411.10329v1

Refer to caption

Dataset (Number of Inappropriate Contents)		ER (%) $\uparrow$
Dataset (Number of Inappropriate Contents)		SD-v2.1	LG	SC	NP	SLD	ESD	CA	Safe-CLIP	SafeGen	Ours
I2P (Handcrafted Prompt)	Hate (91)	-2.20	40.66	-	60.44	52.75	54.95	-2.20	49.45	-	70.33
	Harassment (266)	-5.26	35.34	-	59.40	48.50	42.86	6.77	33.83	-	75.94
	Violence (319)	11.29	35.42	-	52.66	49.84	28.84	9.09	43.57	-	84.01
	Self-harm (309)	2.91	33.66	-	58.90	59.87	42.39	6.15	46.60	-	84.14
	Sexual (674)	33.98	34.27	65.28	48.96	45.25	72.70	16.02	33.23	43.76	93.03
	Shocking (423)	2.84	39.72	-	51.06	46.10	40.90	6.38	43.50	-	85.11
	Illegal activity (255)	7.06	38.82	-	64.31	52.55	46.67	6.27	45.88	-	80.78
	Overall	7.23	36.84	-	56.53	50.69	47.04	6.93	42.29	-	81.91
Adversarial Prompt	SP (581)	25.13	39.07	64.37	56.11	49.05	70.57	16.52	38.38	44.23	91.22
	RAB (3596)	30.34	49.78	96.77	10.17	3.89	31.89	2.50	42.88	81.23	98.92
	MMA (1738)	70.25	77.67	65.88	25.43	9.78	76.01	16.46	79.86	94.53	94.99
	Overall	41.91	55.51	75.67	30.57	20.91	59.49	11.83	53.71	73.33	95.04