notesum.ai

Published at November 4

Adaptive Length Image Tokenization via Recurrent Allocation

cs.CV

cs.AI

cs.LG

cs.RO

Released Date: November 4, 2024

Authors: Shivam Duggal¹, Phillip Isola¹, Antonio Torralba¹, William T. Freeman¹

Aff.: ¹MIT CSAIL

Arxiv: http://arxiv.org/abs/2411.02393v1

Refer to caption

Approach	ImageNet100								COCO			Wikipedia (WIT)
Approach	32	64	96	128	160	192	224	256	$32^{\text{$\#$}}$ / 64	128	256	$32^{\text{$\#$}}$ / 64	128	256
Titok-L-32	11.60	-	-	-	-	-	-	-	$14.18^{\text{$\#$}}$	-	-	$53.57^{\text{$\#$}}$	-	-
Titok-B-64	-	8.22	-	-	-	-	-	-	9.15	-	-	42.86	-	-
Titok-S-128	-	-	-	8.22	-	-	-	-	-	9.15	-	-	38.16	-
VQ-GAN	-	-	-	-	-	-	-	7.04	-	-	7.77	-	-	31.27
Ours-S	22.57	16.17	13.30	11.69	10.22	9.30	8.55	8.25	22.28	14.22	9.72	61.77	47.91	38.45
Ours-SemiLarge	19.70	13.92	11.39	10.41	9.23	8.75	8.22	8.03	-	-	-	-	-	-