notesum.ai

Published at November 26

DreamMix: Decoupling Object Attributes for Enhanced Editability in Customized Image Inpainting

cs.CV

Released Date: November 26, 2024

Authors: Yicheng Yang¹, Pengxiang Li¹, Lu Zhang¹, Liqian Ma², Ping Hu³, Siyu Du¹, Yunzhi Zhuge¹, Xu Jia¹, Huchuan Lu¹

Aff.: ¹Dalian University of Technology; ²ZMO AI; ³University of Electronic Science and Technology of China

Arxiv: http://arxiv.org/abs/2411.17223v1

[Uncaptioned image]

	Method	Identity Preservation				Attribute Editing
	Method	CLIP-T↑	CLIP-I↑	DINO↑	FID↓	CLIP-T↑	CLIP-I↑	DINO↑	FID↓
Large-scale Training	Paint-by-Example [39]	0.238	0.541	0.582	44.6	-	-	-	-
	MimicBrush [5]	0.256	0.614	0.631	47.0	-	-	-	-
	AnyDoor [6]	0.284	0.688	0.711	53.7	-	-	-	-
	IP-Adapter [40]	0.268	0.633	0.633	45.3	0.236	0.611	0.629	47.5
	LAR-Gen [26]	0.269	0.662	0.676	48.7	0.257	0.656	0.668	49.6
Tuning-free	TIGIC [18]	0.262	0.533	0.584	45.2	0.236	0.479	0.548	47.3
Few-shot Fine-tuning	DreamBooth^†	0.248	0.609	0.604	40.8	0.235	0.567	0.595	42.0
	DreamBooth [33]	0.253	0.644	0.628	40.5	0.235	0.601	0.612	41.7
	DreamMix^†	0.284	0.713	0.685	43.4	0.284	0.659	0.672	44.2
	DreamMix	0.285	0.728	0.712	42.9	0.289	0.674	0.695	43.9
	Real Image	0.315	0.883	0.865	35.9	-	-	-	-