notesum.ai

Published at December 10

TraSCE: Trajectory Steering for Concept Erasure

cs.CV

cs.AI

cs.LG

Released Date: December 10, 2024

Authors: Anubhav Jain¹, Yuya Kobayashi², Takashi Shibuya², Yuhta Takida², Nasir Memon¹, Julian Togelius¹, Yuki Mitsufuji³

Aff.: ¹New York University; ²Sony AI; ³Sony Group Corporation

Arxiv: http://arxiv.org/pdf/2412.07658v1

Refer to caption

Method	Ring-A-Bell			MMA-Diffusion $\downarrow$	P4D $\downarrow$	UnLearnDiffAtk $\downarrow$	I2P $\downarrow$	COCO
Method	K77 $\downarrow$	K38 $\downarrow$	K16 $\downarrow$	MMA-Diffusion $\downarrow$	P4D $\downarrow$	UnLearnDiffAtk $\downarrow$	I2P $\downarrow$	FID $\downarrow$	CLIP $\uparrow$
SDv1.4	85.26	87.37	93.68	95.7	98.7	69.7	17.8	16.71	0.304
SA [9]	63.15	56.84	56.84	9.30	47.68	12.68	2.81	25.80	0.297
CA [13]	86.32	91.58	89.47	9.90	10.60	5.63	1.04	24.12	0.301
ESD [5]	20.00	29.47	35.79	12.70	9.27	15.49	2.87	18.18	0.302
MACE [15]	2.10	0.0	0.0	0.50	2.72	2.82	1.51	16.80	0.287
UCE [6]	10.52	9.47	12.61	17.70	29.93	9.86	0.87	17.99	0.302
RECE [7]	5.26	4.21	5.26	13.30	21.77	5.63	0.72	17.74	0.302
SLD-Max [25]	23.16	32.63	42.11	52.10	35.76	9.15	1.74	28.75	0.284
SLD-Strong [25]	56.84	64.21	61.05	61.30	49.01	18.31	2.28	24.40	0.291
SLD-Medium [25]	92.63	88.42	91.58	65.70	68.21	33.10	3.95	21.17	0.298
SD-NP	17.89	28.42	34.74	44.44	24.00	7.80	0.74	18.33	0.301
SAFREE [29]	35.78	47.36	55.78	36.40	40.82	10.56	1.45	19.32	0.301
TraSCE (ours)	1.05	2.10	2.10	16.60	3.97	0.70	0.45	17.41	0.299