notesum.ai

Published at November 15

The Unreasonable Effectiveness of Guidance for Diffusion Models

cs.CV

cs.AI

cs.LG

Released Date: November 15, 2024

Authors: Tim Kaiser¹, Nikolas Adaloglou¹, Markus Kollmann¹

Aff.: ¹Heinrich Heine University of Dusseldorf

Arxiv: http://arxiv.org/abs/2411.10257v1

Refer to caption

	EDM			EDM2-S
	CIFAR-10	CIFAR-100	FFHQ-64	ImageNet-512
	FDD	FDD	FDD	FDD	FID
No Guidance	$139.4$	$196.6$	$174.8$	$112.2$	$2.92$
CFG [16]	$99.9_{0.6}$	$126.6_{0.6}$	$-$	$52.9_{0.8}^{\dagger}$	$2.44_{0.1}^{\dagger}$
Weak model guidance (WMG) methods
- Reduced training	$60.8_{1.1}$	$81.2_{1.3}$	$94.0_{1.0}$	$46.5_{1.2}$	$1.79_{0.8}$
- Reduced capacity	$62.4_{0.9}$	$93.8_{1.1}$	$102.5_{1.3}$	$70.0_{1.4}$	$2.20_{0.7}$
- Reduced capacity and training (RCT)	$62.1_{0.8}$	$87.0_{1.0}$	$95.3_{1.0}$	$\mathbf{42.1_{1.4}}$	$\mathbf{1.67}_{0.8}$
- Weight decay fine-tuning	$65.0_{1.3}$	$88.1_{1.4}$	$111.2_{2.4}$	$52.1_{0.6}$	$2.46_{0.1}$
- Weight decay re-training	$64.7_{1.3}$	$93.3_{1.3}$	$133.0_{1.4}$	$-$	$-$
- Weight decay and reduced training	$\mathbf{59.0_{0.9}}$	$\mathbf{75.9_{1.1}}$	$\mathbf{92.6_{1.0}}$	$44.7_{0.8}$	$2.26_{0.2}$