notesum.ai

Published at November 21

Safety Without Semantic Disruptions: Editing-free Safe Image Generation via Context-preserving Dual Latent Reconstruction

cs.CV

cs.AI

Released Date: November 21, 2024

Authors: Jordan Vice¹, Naveed Akhtar¹, Richard Hartley², Ajmal Mian³

Aff.: ¹University of Melbourne; ²Australian National University; ³University of Western Australia

Arxiv: http://arxiv.org/abs/2411.13982v1

Refer to caption

	I2P [45]								Semantic Disruption
Model (+ Edit)	Hate	Harassment	Violence	Self-harm	Sexual	Shocking	Illegal Act.	Avg.	Edited?	Avg._R	Avg._P
SD2.1	42.7	39.0	41.9	42.0	26.5	51.6	37.7	36.9	✗	0.0	0.0
+ SafeCLIP [36]	25.5	20.7	21.6	16.7	11.8	23.7	16.2	17.2	✓	16.6	16.5
+ SLD-Weak [46]	31.3	28.3	29.7	26.8	14.0	37.6	26.9	25.3	✗	0.0	0.0
+ SLD-Medium [46]	24.5	22.2	22.3	15.7	8.30	26.4	17.3	17.4	✗	0.0	0.0
+ SLD-Strong [46]	19.7	17.4	17.4	8.50	5.60	19.1	11.9	12.4	✗	0.0	0.0
+ Ours @ $\mathbf{w}_{\tilde{\mathbf{x}_{i}}}=0.75$	27.0	26.2	32.1	24.3	35.9	30.8	26.7	29.4	✗	0.0	0.0
+ Ours @ $\mathbf{w}_{\tilde{\mathbf{x}_{i}}}=0.85$	21.1	28.7	25.1	17.6	24.1	17.2	17.8	21.9	✗	0.0	0.0
+ Ours @ $\mathbf{w}_{\tilde{\mathbf{x}_{i}}}=0.95$	21.5	34.0	9.03	2.34	10.3	1.22	10.2	11.9	✗	0.0	0.0
SD1.4	38.6	38.8	44.9	48.6	64.8	61.2	38.3	48.9	✗	0.0	0.0
+ SafeCLIP [36]	24.1	22.3	31.5	31.6	45.6	38.6	23.0	32.6	✓	16.6	17.5
+ UCE_(⋆) [15]	36.4	29.5	34.1	30.8	25.5	41.1	29.0	31.3	✓	18.3	13.5
+ Receler [23]	28.6	21.7	27.1	24.8	29.4	34.8	21.3	27.0	✓	12.8	13.8
+ SLD-Weak [46]	30.6	24.1	32.1	27.8	13.9	41.9	25.7	25.6	✗	0.0	0.0
+ SLD-Medium [46]	21.6	17.5	23.7	17.4	8.9	31.2	16.7	17.7	✗	0.0	0.0
+ SLD-Strong [46]	15.9	13.6	18.8	11.1	7.8	21.5	11.2	13.5	✗	0.0	0.0
+ Self-Disc. [28]	29.0	18.0	30.0	28.0	22.0	36.0	23.0	27.0	✗	0.0	0.0
+ Ours @ $\mathbf{w}_{\tilde{\mathbf{x}_{i}}}=0.75$	27.1	28.0	35.0	34.1	43.4	45.3	28.6	35.5	✗	0.0	0.0
+ Ours @ $\mathbf{w}_{\tilde{\mathbf{x}_{i}}}=0.85$	16.1	30.1	32.9	35.3	31.1	37.0	29.2	31.9	✗	0.0	0.0
+ Ours @ $\mathbf{w}_{\tilde{\mathbf{x}_{i}}}=0.95$	10.4	12.5	15.7	5.05	18.0	2.66	18.5	12.8	✗	0.0	0.0