notesum.ai

Published at November 11

OmniEdit: Building Image Editing Generalist Models Through Specialist Supervision

cs.CV

cs.AI

Released Date: November 11, 2024

Authors: Cong Wei¹, Zheyang Xiong², Weiming Ren³, Xinrun Du⁴, Ge Zhang¹, Wenhu Chen¹

Aff.: ¹University of Waterloo; ²University of Wisconsin-Madison; ³Vector Institute; ⁴M-A-P

Arxiv: http://arxiv.org/abs/2411.07199v1

Refer to caption

Models	VIEScore (GPT4o)			VIEScore (Gemini)			Human Evaluation
	$PQ_{avg}\uparrow$	$SC_{avg}\uparrow$	$O_{avg}\uparrow$	$PQ_{avg}\uparrow$	$SC_{avg}\uparrow$	$O_{avg}\uparrow$	$PQ_{avg}\uparrow$	$SC_{avg}\uparrow$	$O_{avg}\uparrow$	$Acc_{avg}\uparrow$
Inversion-based Methods
DiffEdit	5.88	2.73	2.79	6.09	2.01	2.39	-	-	-	-
SDEdit	6.71	2.18	2.78	6.31	2.06	2.48	-	-	-	-
End-to-End Methods
InstructPix2Pix	7.05	3.04	3.45	6.46	1.88	2.31	-	-	-	-
MagicBrush	6.11	3.53	3.60	6.36	2.27	2.61	-	-	-	-
UltraEdit(SD-3)	6.44	4.66	4.86	6.49	4.33	4.45	0.72	0.52	0.57	0.20
HQ-Edit	5.42	2.15	2.25	6.18	1.71	1.96	0.80	0.27	0.29	0.10
CosXL-Edit	8.34	5.81	6.00	7.01	4.90	4.81	0.82	0.56	0.59	0.35
HIVE	5.35	3.65	3.57	5.84	2.84	3.05	-	-	-	-
Omni-Edit	8.38	6.66	6.98	7.06	5.82	5.78	0.83	0.71	0.69	0.55
$\Delta$ - Best baseline	+0.04	+0.85	+0.98	+0.05	+0.92	+0.97	+0.01	+0.15	+0.10	+0.20