Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Lower-neck involvement and post-concurrent chemoradiotherapy response predict distant metastasis in head and neck cancer.

Radiation oncology journal·2026
Same author

Generation of Functional Oligodendrocyte Progenitor Cells Through Serial Replating of iPSC-Derived NPC Spheres.

Cells·2026
Same author

Habitual Ultra-processed Food Intake Is Associated with Gut Dysbiosis and Pro-inflammatory Metabolite Profiles in Korean Patients with IBD.

Digestive diseases and sciences·2026
Same author

Population Pharmacokinetics and Model-Informed Precision Dosing of Lenalidomide Incorporating Total and Unbound Plasma Concentrations in Renally Impaired Patients with Multiple Myeloma.

Drug design, development and therapy·2026
Same author

Advancing Small-Molecule Immunotherapy Through Polymeric Micelle Delivery.

Pharmaceutics·2026
Same author

Korean Medical Consultation With Open-Weight Large Language Models: Pilot Comparative Evaluation of Retrieval-Augmented Generation With Metadata Filtering.

JMIR formative research·2026

Related Experiment Video

Updated: Sep 13, 2025

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
03:14

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Published on: December 6, 2024

684

Visual delta generation with large multi-modal models enhances composed image retrieval using unlabeled data.

Young Kyun Jang1, Donghyun Kim2

  • 1Meta Platforms (United States), Menlo Park, USA.

Scientific Reports
|July 28, 2025
PubMed
Summary
This summary is machine-generated.

This study introduces a new method for Composed Image Retrieval (CIR) using unlabeled data. The Visual Delta Generator (VDG) creates pseudo-triplets to improve CIR performance in various settings, achieving state-of-the-art results.

More Related Videos

Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
08:25

Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment

Published on: May 7, 2019

9.1K
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
04:48

Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography

Published on: November 30, 2022

2.9K

Related Experiment Videos

Last Updated: Sep 13, 2025

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
03:14

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Published on: December 6, 2024

684
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
08:25

Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment

Published on: May 7, 2019

9.1K
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
04:48

Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography

Published on: November 30, 2022

2.9K

Area of Science:

  • Computer Science
  • Artificial Intelligence
  • Machine Learning

Background:

  • Composed Image Retrieval (CIR) typically relies on supervised learning with labeled triplets, which are difficult to obtain and limit scalability.
  • Weakly supervised (zero-shot) CIR uses image-caption pairs but often yields lower accuracy and cannot be directly applied to unlabeled data scenarios.
  • Existing CIR methods struggle to leverage fully unlabeled data for tasks like semi-supervised learning, domain adaptation, and test-time adaptation.

Purpose of the Study:

  • To extend Composed Image Retrieval (CIR) applications to semi-supervised learning, domain adaptation, and test-time adaptation using only unlabeled image data.
  • To propose a novel approach that overcomes the limitations of existing CIR methods in handling unlabeled data.
  • To enhance CIR performance across diverse settings by generating useful training signals from unlabeled image collections.

Main Methods:

  • Developed a Visual Delta Generator (VDG) utilizing a large language model to describe visual differences between reference and target images.
  • Trained VDG to generate textual descriptions of visual deltas, creating pseudo-triplets from unlabeled auxiliary image data.
  • Employed these pseudo-triplets to boost the performance of CIR models in semi-supervised, domain adaptation, and test-time adaptation contexts.

Main Results:

  • The proposed VDG approach significantly improves existing supervised learning methods for CIR.
  • Achieved state-of-the-art results on established Composed Image Retrieval benchmarks.
  • Demonstrated the effectiveness of the method in expanding CIR applications to semi-supervised, domain adaptation, and test-time adaptation settings.

Conclusions:

  • The VDG method offers a scalable and effective solution for Composed Image Retrieval using unlabeled data.
  • This approach broadens the applicability of CIR to new domains and adaptation scenarios without requiring labeled triplets.
  • The model-agnostic nature of VDG allows for seamless integration and performance enhancement of various CIR models.