notesum.ai

Published at October 31

Disentangling Disentangled Representations: Towards Improved Latent Units via Diffusion Models

cs.LG

cs.AI

cs.CV

Released Date: October 31, 2024

Authors: Youngjun Jun¹, Jiwoo Park¹, Kyobin Choo¹, Tae Eun Choi¹, Seong Jae Hwang¹

Aff.: ¹Yonsei University

Arxiv: http://arxiv.org/abs/2410.23820v1

Refer to caption

	Method	Cars3D		Shapes3D		Mpi3D-toy
	Method	FactorVAE score $\uparrow$	DCI $\uparrow$	FactorVAE score $\uparrow$	DCI $\uparrow$	FactorVAE score $\uparrow$	DCI $\uparrow$
VAE	FactorVAE [31]	$0.906\pm 0.052$	$0.161\pm 0.019$	$0.840\pm 0.066$	$0.611\pm 0.082$	$0.152\pm 0.025$	$0.240\pm 0.051$
	$\beta$ -TCVAE [9]	$0.855\pm 0.082$	$0.140\pm 0.019$	$0.873\pm 0.074$	$0.613\pm 0.114$	$0.179\pm 0.017$	$0.237\pm 0.056$
	DAVA [16]	$0.940\pm 0.010$	$0.230\pm 0.040$	$0.820\pm 0.030$	$0.780\pm 0.030$	$0.480\pm 0.050$	$0.270\pm 0.030$
GAN	ClosedForm [54]	$0.873\pm 0.036$	$0.243\pm 0.048$	$0.951\pm 0.021$	$0.525\pm 0.078$	$0.523\pm 0.056$	$0.318\pm 0.014$
	GANSpace [19]	$0.932\pm 0.018$	$0.209\pm 0.031$	$0.788\pm 0.091$	$0.284\pm 0.034$	$0.465\pm 0.036$	$0.229\pm 0.042$
	DisCo-GAN [48]	$0.855\pm 0.074$	$0.271\pm 0.037$	$0.877\pm 0.031$	$0.708\pm 0.048$	$0.371\pm 0.030$	$0.292\pm 0.024$
DM	DisDiff-VQ [69]	$\bm{0.976}\pm 0.018$	$0.232\pm 0.019$	$0.902\pm 0.043$	$0.723\pm 0.013$	$0.617\pm 0.070$	$0.337\pm 0.057$
	FDAE [66]	$0.918\pm 0.027$	$0.232\pm 0.418$	$0.987\pm 0.023$	$0.917\pm 0.038$	-	-
	EncDiff [67]	$0.773\pm 0.060$	$\underline{0.279}\pm 0.022$	$\underline{0.999}\pm 0.000$	$\bm{0.969}\pm 0.030$	$\underline{0.872}\pm 0.049$	$\bm{0.685}\pm 0.044$
	Ours	$\underline{0.941}\pm 0.002$	$\bm{0.414}\pm 0.013$	$\bm{1.000}\pm 0.000$	$\underline{0.938}\pm 0.001$	$\bm{0.930}\pm 0.004$	$\underline{0.627}\pm 0.002$