notesum.ai

Published at November 6

DiMSUM: Diffusion Mamba -- A Scalable and Unified Spatial-Frequency Method for Image Generation

cs.CV

cs.AI

Released Date: November 6, 2024

Authors: Hao Phung¹, Quan Dao², Trung Dao³, Hoang Phan⁴, Dimitris Metaxas, Anh Tran³

Aff.: ¹Cornell University; ²Rutgers University; ³VinAI Research; ⁴New York University

Arxiv: http://arxiv.org/abs/2411.04168v1

[Uncaptioned image]

Model	NFE $\downarrow$	FID $\downarrow$	Recall $\uparrow$	Epochs
Ours	61	4.62	0.52	225
Zigma† [25]	65	7.66	0.40	400
LFM-8 [5]	89	5.26	0.46	500
LDM-4 [51] (ADM)	500	5.11	0.49	600
LDM-8 (ADM)‡	250	15.37	-	500
LDM-8 (DiT)‡	250	10.21	-	500
LSGM [60]	23	7.22	-	1K
WaveDiff [49]	2	5.94	0.37	500
DDGAN [64]	2	7.64	0.36	800
RDUOT [6]	2	5.60	0.38	-
Score SDE [58]	4000	7.23	-	6.2K