ML-CLIPSim: Multi-Layer CLIP Similarity for Machine-Oriented Image Quality

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ding, Feng, Fu, Haisheng, Liang, Jie, Xu, Qihan, Zhu, Siyu, Han, Jingning
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913108174307328
author Ding, Feng
Fu, Haisheng
Liang, Jie
Xu, Qihan
Zhu, Siyu
Han, Jingning
author_facet Ding, Feng
Fu, Haisheng
Liang, Jie
Xu, Qihan
Zhu, Siyu
Han, Jingning
contents We study full-reference image quality assessment from a machine-centric perspective, where images are evaluated by how well they preserve information for downstream models. We formulate machine-oriented quality as a latent machine utility and approximate it through pairwise predictive-consistency comparisons. To this end, we construct PCMP, a dataset of PSNR-matched distortion pairs labeled by consistency votes from multiple pretrained models. We further propose ML-CLIPSim, a differentiable quality metric built on a frozen CLIP visual encoder, which aggregates intermediate patch-token similarities and global image embeddings. Experiments on machine-preference benchmarks, human-IQA datasets, and learned image compression show that ML-CLIPSim better aligns with machine-oriented preferences than conventional fidelity and perceptual metrics, while remaining competitive for human quality prediction. Used as a compression distortion term, it improves rate--task trade-offs across multiple downstream tasks.
format Preprint
id arxiv_https___arxiv_org_abs_2605_09479
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle ML-CLIPSim: Multi-Layer CLIP Similarity for Machine-Oriented Image Quality
Ding, Feng
Fu, Haisheng
Liang, Jie
Xu, Qihan
Zhu, Siyu
Han, Jingning
Image and Video Processing
Computer Vision and Pattern Recognition
Multimedia
We study full-reference image quality assessment from a machine-centric perspective, where images are evaluated by how well they preserve information for downstream models. We formulate machine-oriented quality as a latent machine utility and approximate it through pairwise predictive-consistency comparisons. To this end, we construct PCMP, a dataset of PSNR-matched distortion pairs labeled by consistency votes from multiple pretrained models. We further propose ML-CLIPSim, a differentiable quality metric built on a frozen CLIP visual encoder, which aggregates intermediate patch-token similarities and global image embeddings. Experiments on machine-preference benchmarks, human-IQA datasets, and learned image compression show that ML-CLIPSim better aligns with machine-oriented preferences than conventional fidelity and perceptual metrics, while remaining competitive for human quality prediction. Used as a compression distortion term, it improves rate--task trade-offs across multiple downstream tasks.
title ML-CLIPSim: Multi-Layer CLIP Similarity for Machine-Oriented Image Quality
topic Image and Video Processing
Computer Vision and Pattern Recognition
Multimedia
url https://arxiv.org/abs/2605.09479