notesum.ai

Published at November 29

JetFormer: An Autoregressive Generative Model of Raw Images and Text

cs.LG

cs.AI

cs.CV

Released Date: November 29, 2024

Authors: Michael Tschannen¹, André Susano Pinto, Alexander Kolesnikov¹

Aff.: ¹Google DeepMind

Arxiv: http://arxiv.org/pdf/2411.19722v1

Refer to caption

	extra step	FID	Precision	Recall	NLL
BigGAN-deep (Brock et al., 2018)	–	6.95	0.87	0.28
ADM-G (Dhariwal & Nichol, 2021)	–	4.59	0.82	0.52
LDM-4-G (Rombach et al., 2022)	VAE	3.60	0.87	0.48
VQGAN (Esser et al., 2020)	VQ-VAE	5.20
ViT-VQGAN (Yu et al., 2022a)	VQ-VAE	3.04
GIVT-Causal (Tschannen et al., 2024)	VAE	3.35	0.84	0.53
JetFormer-B	–	7.25	0.72	0.44	3.06
JetFormer-L	–	6.64	0.69	0.56	3.05