notesum.ai

Published at November 2

cs.CV

cs.AI

cs.GR

cs.LG

Released Date: November 2, 2024

Authors: Wonguk Cho¹, Seokeon Choi¹, Debasmit Das¹, Matthias Reisser¹, Taesup Kim², Sungrack Yun¹, Fatih Porikli¹

Aff.: ¹Qualcomm AI Research; ²Seoul National University

Method	# of Parameters		Training Memory		DreamBooth			CustomConcept101
Method	Base	LoRA	Peak	Comp. w/ Inf.	DINO	CLIP-I	CLIP-T	DINO	CLIP-I	CLIP-T
Full FT	866M	-	16.62GB	+376%	0.663 $\pm 0.013$	0.802 $\pm 0.007$	0.302 $\pm 0.002$	0.605 $\pm 0.005$	0.773 $\pm 0.006$	0.302 $\pm 0.002$
LoRA FT (r=128)	866M	27M	5.23GB	+50%	0.658 $\pm 0.001$	0.806 $\pm 0.005$	0.299 $\pm 0.002$	0.603 $\pm 0.008$	0.773 $\pm 0.005$	0.302 $\pm 0.002$
LoRA FT (r=1)	866M	207K	4.84GB	+39%	0.516 $\pm 0.011$	0.738 $\pm 0.003$	0.314 $\pm 0.001$	0.522 $\pm 0.008$	0.737 $\pm 0.005$	0.305 $\pm 0.001$
Hollowed Net (Ours)	527M	24M	3.88GB	+11%	0.660 $\pm 0.011$	0.805 $\pm 0.006$	0.300 $\pm 0.001$	0.603 $\pm 0.007$	0.773 $\pm 0.005$	0.302 $\pm 0.002$