Model Cards for IQA-PyTorch

General FR/NR Methods

List all model names with:

import pyiqa
print(pyiqa.list_models())

FR Method	Model names	Description	Score Direction
DMM	dmm	Published in TMM	Higher is better
TOPIQ	topiq_fr, topiq_fr-pipal	Proposed in this paper	Higher is better
AHIQ	ahiq		Higher is better
PieAPP	pieapp		Lower is better
LPIPS	lpips, lpips-vgg, stlpips, stlpips-vgg, lpips+, lpips-vgg+		Lower is better
DISTS	dists		Lower is better
WaDIQaM	wadiqam_fr		Higher is better
CKDN1	ckdn		Higher is better
FSIM	fsim		Higher is better
SSIM	ssim, ssimc	Gray input (y channel), color input	Higher is better
MS-SSIM	ms_ssim		Higher is better
CW-SSIM	cw_ssim		Higher is better
PSNR	psnr, psnry	Color input, gray input (y channel)	Higher is better
VIF	vif		Higher is better
GMSD	gmsd		Lower is better
NLPD	nlpd		Lower is better
VSI	vsi		Higher is better
MAD	mad		Lower is better

NR Method	Model names	Description	Score Direction
Q-Align	qalign (with quality[default], aesthetic options)	Large vision-language models	Higher is better
QualiCLIP(+)	qualiclip, qualiclip+, qualiclip+-clive, qualiclip+-flive, qualiclip+-spaq	QualiCLIP(+) with different datasets, koniq by default	Higher is better
MACLIP	maclip	CLIP based method	Higher is better
LIQE	liqe, liqe_mix	CLIP based method	Higher is better
ARNIQA	arniqa, arniqa-live, arniqa-csiq, arniqa-tid, arniqa-kadid, arniqa-clive, arniqa-flive, arniqa-spaq	ARNIQA with different datasets, koniq by default	Higher is better
TOPIQ	topiq_nr, topiq_nr-flive, topiq_nr-spaq	TOPIQ with different datasets, koniq by default	Higher is better
TReS	tres, tres-flive	TReS with different datasets, koniq by default	Higher is better
FID	fid	Statistic distance between two datasets	Lower is better
CLIPIQA(+)	clipiqa, clipiqa+, clipiqa+_vitL14_512,clipiqa+_rn50_512	CLIPIQA(+) with different backbone, RN50 by default	Higher is better
MANIQA	maniqa, maniqa-kadid, maniqa-pipal	MUSIQ with different datasets, koniq by default	Higher is better
MUSIQ	musiq, musiq-spaq, musiq-paq2piq, musiq-ava	MUSIQ with different datasets, koniq by default	Higher is better
DBCNN	dbcnn		Higher is better
PaQ-2-PiQ	paq2piq		Higher is better
HyperIQA	hyperiqa		Higher is better
NIMA	nima, nima-vgg16-ava	Aesthetic metric trained with AVA dataset	Higher is better
WaDIQaM	wadiqam_nr		Higher is better
CNNIQA	cnniqa		Higher is better
NRQM(Ma)2	nrqm	No backward	Higher is better
PI(Perceptual Index)	pi	No backward	Lower is better
BRISQUE	brisque, brisque_matlab	No backward	Lower is better
ILNIQE	ilniqe	No backward	Lower is better
NIQE	niqe, niqe_matlab	No backward	Lower is better
PIQE	piqe	No backward	Lower is better

[1] This method use distorted image as reference. Please refer to the paper for details.
[2] Currently, only naive random forest regression is implemented and does not support backward.

IQA Methods for Specific Tasks

Task	Method	Description	Score Direction
Color IQA	msswd	Perceptual color difference metric MS-SWD, ECCV2024, Arxiv, Github	Lower is better
Face IQA	topiq_nr-face	TOPIQ model trained with face IQA dataset (GFIQA)	Higher is better
Underwater IQA	uranker	A ranking-based underwater image quality assessment (UIQA) method, AAAI2023, Arxiv, Github	Higher is better

Metric Output Score Range

Note: ~ means that the corresponding numeric bound is typical value and not mathematically guaranteed

You can now access the rough output range of each metric like this:

metric = pyiqa.create_metric('lpips')
print(metric.score_range)