notesum.ai

Published at November 19

PoM: Efficient Image and Video Generation with the Polynomial Mixer

cs.CV

cs.AI

Released Date: November 19, 2024

Authors: David Picard¹, Nicolas Dufour²

Aff.: ¹LIGM, Ecole Nationale des Ponts et Chaussées, IP Paris, Univ Gustave Eiffel, CNRS, France; ²LIGM, Ecole Nationale des Ponts et Chaussées, IP Paris, Univ Gustave Eiffel, CNRS, France; LIX, Ecole Polytechnique, IP Paris, CNRS, France

Arxiv: http://arxiv.org/abs/2411.12663v1

Model	Sample config	#train	FID $\downarrow$	IS $\uparrow$	Precision $\uparrow$	Recall $\uparrow$
Mask-GIT [8]			6.18	182.1	0.80	0.51
DIFFUSSM-XL^† [79]	250 steps DDPM	660M	2.28	259.1	0.86	0.56
DiM-H^† [69]	25 steps DPM++	480M	2.21	-	-	-
ADM-G [16]	250 steps DDIM	500M	4.59	186.7	0.83	0.53
LDM-4-G [63]	250 steps DDIM	215M	3.60	247.7	0.87	0.48
RIN [40]	1000 steps DDPM	600M	3.42	182.0	-	-
DiT-XL/2^† [58]	250 steps DDPM	1.8B	2.27	278.2	0.83	0.57
SiT-XL/2^† [56]	125 steps Heun	1.8B	2.15	254.9	0.81	0.60
DiPoM-XL/2 $\mathcal{L}_{\text{D}}$ (ours)	250 steps DDIM	950M	2.46	240.6	0.78	0.60
DiPoM-XL/2 $\mathcal{L}_{\text{FM}}$ (ours)	125 steps Heun	950M	3.70	255.2	0.79	0.56