Session 24: Advancing Image Classification with Discrete Diffusion Classification Modeling
Diffusion LLM
Hello folks,
Image classification is a well-studied task in computer vision, yet it remains challenging under high-uncertainty conditions, such as when input images are corrupted or training data are limited. Conventional classification approaches typically train models to directly predict class labels from input images, but this might lead to suboptimal performance in such scenarios.
To address this issue, the authors propose Discrete Diffusion Classification Modeling (DiDiCM), a novel framework that leverages a diffusion-based procedure to model the posterior distribution of class labels conditioned on the input image.
DiDiCM supports diffusion-based predictions either on class probabilities or on discrete class labels, providing flexibility in computation and memory trade-offs.
They conduct a comprehensive empirical study demonstrating the superior performance of DiDiCM over standard classifiers, showing that a few diffusion iterations achieve higher classification accuracy on the ImageNet dataset compared to baselines, with accuracy gains increasing as the task becomes more challenging.
This Monday, Omer Belhasin from Technion and NVIDIA will present his work: Advancing Image Classification with Discrete Diffusion Classification Modeling.
Title: Advancing Image Classification with Discrete Diffusion Classification Modeling
Meeting Link: click here
Time: July 27 (Monday) 1pm ET / 10am PT / 7pm CET / 10:30pm IST
Paper: [2511.20263] Advancing Image Classification with Discrete Diffusion Classification Modeling
Prior knowledge:
Fundamentals of discrete diffusion (video by Sasha Rush)
The Diffusion Duality (video by our reading group)
Abstract:
Image classification is a well-studied task in computer vision, and yet it remains challenging under high-uncertainty conditions, such as when input images are corrupted or training data are limited. Conventional classification approaches typically train models to directly predict class labels from input images, but this might lead to suboptimal performance in such scenarios. To address this issue, we propose Discrete Diffusion Classification Modeling (DiDiCM), a novel framework that leverages a diffusion-based procedure to model the posterior distribution of class labels conditioned on the input image. DiDiCM supports diffusion-based predictions either on class probabilities or on discrete class labels, providing flexibility in computation and memory trade-offs. We conduct a comprehensive empirical study demonstrating the superior performance of DiDiCM over standard classifiers, showing that a few diffusion iterations achieve higher classification accuracy on the ImageNet dataset compared to baselines, with accuracy gains increasing as the task becomes more challenging.
Yours truly,
Subham, Justin, Zhihan
Diffusion LLM
Gentle reminder: See you all at 1pm ET / 10am PT / 7pm CET / 10:30pm IST
Meeting Link: click here
Today's paper: [2511.20263] Advancing Image Classification with Discrete Diffusion Classification Modeling