Semi-Supervised Conditional Diffusion via Label Augmentation

2026-07-18

AI Summary

This work proposes Label-Augmented Conditional Diffusion (LACD), a novel approach for efficient semi-supervised conditional generation in label-scarce settings. By introducing dedicated dummy labels for unlabeled samples and performing joint denoising score matching on the augmented dataset, LACD enables effective utilization of limited labeled data. The study establishes, for the first time, identifiability conditions for conditional diffusion models under semi-supervised learning and rigorously proves that LACD achieves a strictly faster convergence rate in total variation distance compared to purely supervised counterparts. Empirical evaluations across synthetic, image, and tabular benchmarks demonstrate that LACD substantially improves both sample efficiency and generation quality, confirming its theoretical advantages and practical efficacy.

Abstract

Conditional diffusion models have become a powerful and flexible framework for learning complex conditional distributions from labeled data. In practice, however, acquiring high-quality labels is costly and time-consuming, leaving large volumes of unlabeled data unused. To address this, we introduce label-augmented conditional diffusion (LACD), a simple and effective approach that incorporates unlabeled examples by assigning them a designated trivial label and performing joint denoising score matching over the augmented dataset. We provide sufficient conditions guaranteeing population-level identifiability of the target conditional distribution under this scheme. Moreover, we establish rigorous statistical guarantees: when sufficiently many unlabeled samples are available, the sampling distribution produced by LACD converges strictly faster than the purely supervised estimator in total variation distance, and at least as fast in Wasserstein-1 distance. Extensive experiments on synthetic, image, and tabular benchmarks corroborate our theory and show substantial gains in sample efficiency and generative performance compared with the purely supervised estimator.

Problem

• semi-supervised learning
• conditional diffusion
• label scarcity
• unlabeled data
• sample efficiency

Innovation

• semi-supervised learning
• conditional diffusion models
• label augmentation
• score matching
• statistical convergence