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 Experiment Video

Updated: Jul 12, 2026

Quantifying Intermembrane Distances with Serial Image Dilations
07:45

Quantifying Intermembrane Distances with Serial Image Dilations

Published on: September 28, 2018

ELBO-T2IAlign: A Generic ELBO-Based Method for Calibrating Pixel-Level Text-Image Alignment in Diffusion Models.

Qin Zhou, Zhiyang Zhang, Jinglong Wang

    IEEE Transactions on Image Processing : a Publication of the IEEE Signal Processing Society
    |July 9, 2026
    PubMed
    Summary

    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

    MUFOLD-DB: a processed protein structure database for protein structure prediction and analysis.

    BMC genomics·2015
    Same author

    The I-TASSER Suite: protein structure and function prediction.

    Nature methods·2014
    Same author

    Genome-wide expression analysis of soybean NF-Y genes reveals potential function in development and drought response.

    Molecular genetics and genomics : MGG·2014
    Same author

    Classification of lung cancer using ensemble-based feature selection and machine learning methods.

    Molecular bioSystems·2014
    Same author

    Resveratrol possesses protective effects in a pristane-induced lupus mouse model.

    PloS one·2014
    Same author

    Protein-losing enteropathy in systemic lupus erythematosus: 12 years experience from a Chinese academic center.

    PloS one·2014

    Diffusion models can generate images but struggle with perfect text-image alignment, especially for small or rare objects. A new training-free method, ELBO-T2IAlign, calibrates this alignment for improved downstream tasks.

    Area of Science:

    • Artificial Intelligence
    • Computer Vision
    • Machine Learning

    Background:

    • Diffusion models generate high-quality images and encode text-image alignment.
    • Current methods assume perfect alignment, which is often not true.
    • Pixel-level and class-level text-image misalignment is a significant issue.

    Purpose of the Study:

    • Evaluate pixel-level image and class-level text alignment in diffusion models.
    • Analyze the causes of misalignment, particularly training data bias.
    • Propose a method to calibrate text-image alignment in diffusion models.

    Main Methods:

    • Utilized zero-shot referring image segmentation as a proxy task.
    • Analyzed misalignment in diffusion models concerning training data bias.

    More Related Videos

    Medical-grade Sterilizable Target for Fluid-immersed Fetoscope Optical Distortion Calibration
    07:03

    Medical-grade Sterilizable Target for Fluid-immersed Fetoscope Optical Distortion Calibration

    Published on: February 23, 2017

    Related Experiment Videos

    Last Updated: Jul 12, 2026

    Quantifying Intermembrane Distances with Serial Image Dilations
    07:45

    Quantifying Intermembrane Distances with Serial Image Dilations

    Published on: September 28, 2018

    Medical-grade Sterilizable Target for Fluid-immersed Fetoscope Optical Distortion Calibration
    07:03

    Medical-grade Sterilizable Target for Fluid-immersed Fetoscope Optical Distortion Calibration

    Published on: February 23, 2017

  • Introduced ELBO-T2IAlign, a training-free calibration method based on evidence lower bound (ELBO).

    • Main Results:

    • Identified misalignment issues in images with small-sized, occluded, or rare object classes.
    • Demonstrated that ELBO-T2IAlign effectively calibrates pixel-text alignment.
    • Showcased improvements in downstream tasks like segmentation and image editing.

    Conclusions:

    • ELBO-T2IAlign is a generic and effective method for improving text-image alignment in diffusion models.
    • The training-free approach requires no additional annotations or retraining.
    • Calibration enhances performance across various complementary downstream applications.