Diffusion-Based Cross-Modal Feature Extraction for Multi-Label Classification

Abstract

Multi-label classification has broad applications and depends on powerful representations capable of capturing multi-label interactions. We introduce \textit{Diff-Feat}, a simple but powerful framework that extracts intermediate features from pre-trained diffusion-Transformer models for images and text, and fuses them for downstream tasks. We observe that for vision tasks, the most discriminative intermediate feature along the diffusion process occurs at the middle step and is located in the middle block in Transformer. In contrast, for language tasks, the best feature occurs at the noise-free step and is located in the deepest block. In particular, we observe a striking phenomenon across varying datasets: a mysterious "Layer $12$" consistently yields the best performance on various downstream classification tasks for images (under DiT-XL/2-256$\times$256). We devise a heuristic local-search algorithm that pinpoints the locally optimal "image-text"$\times$"block-timestep" pair among a few candidates, avoiding an exhaustive grid search. A simple fusion-linear projection followed by addition-of the selected representations yields state-of-the-art performance: 98.6\% mAP on MS-COCO-enhanced and 45.7\% mAP on Visual Genome 500, surpassing strong CNN, graph, and Transformer baselines by a wide margin. t-SNE and clustering metrics further reveal that \textit{Diff-Feat} forms tighter semantic clusters than unimodal counterparts. The code is available at https://github.com/lt-0123/Diff-Feat.