notesum.ai

Published at December 4

Appearance Matching Adapter for Exemplar-based Semantic Image Synthesis

cs.CV

Released Date: December 4, 2024

Authors: Siyoon Jin¹, Jisu Nam², Jiyoung Kim¹, Dahyun Chung¹, Yeong-Seok Kim³, Joonhyung Park³, Heonjeong Chu³, Seungryong Kim²

Aff.: ¹Korea University; ²KAIST; ³Hyundai Mobis

Arxiv: http://arxiv.org/pdf/2412.03150v1

[Uncaptioned image]

Methods	Training	BDD100K [42]			Cityscapes [4]
Methods	Training	Self-sim. $(\downarrow)$	$I_{\mathrm{CLIP}}$ $(\uparrow)$	FID $(\downarrow)$	Self-sim. $(\downarrow)$	$I_{\mathrm{CLIP}}$ $(\uparrow)$	FID $(\downarrow)$
ControlNet [43]	✓	0.052	0.638	95.10	0.058	0.761	113.87
ControlNeXt [29]	✓	0.047	0.731	94.48	0.055	0.760	130.04
FreeControl [23]	✗	0.062	0.575	179.00	0.073	0.690	211.28
Cross-Image Attention [1]	✗	0.177	0.643	233.76	0.133	0.754	214.31
Ctrl-X [18]	✗	0.051	0.700	107.06	0.058	0.818	112.49
ControlNet [43] + IP-Adapter [41]	✓	0.049	0.805	84.82	0.049	0.823	115.98
AM-Adapter (Ours)	✓	0.041	0.819	75.89	0.048	0.835	85.67