LACD diffusion model taps unlabeled data for faster training
Trivial label trick lets diffusion models learn from unlabeled data with guarantees
A new paper from Jin Su, Yuan Gao, Yong Zhou, and Jian Huang tackles a core bottleneck in conditional diffusion models: the scarcity of labeled data. They propose Label-Augmented Conditional Diffusion (LACD), which simply assigns a designated trivial label to unlabeled examples and trains the model via joint denoising score matching over the augmented dataset. The authors provide theoretical guarantees that the target conditional distribution remains identifiable under this scheme, and that the sampling distribution converges strictly faster in total variation distance than a purely supervised estimator when sufficient unlabeled data is available.
Extensive experiments on synthetic data, image datasets, and tabular benchmarks confirm the theory, showing substantial gains in sample efficiency and generative quality without sacrificing distributional fidelity. This work formalizes a practical semi-supervised extension to diffusion models, opening the door to leveraging vast amounts of unlabeled data that are often discarded. The paper (34 pages, 7 figures) is available on arXiv under arXiv:2607.16685.
- Assigns unlabeled data a designated trivial label for joint denoising score matching training
- Proves population-level identifiability and faster convergence in total variation distance with enough unlabeled samples
- Achieves substantial gains in sample efficiency and generative quality on synthetic, image, and tabular benchmarks
Why It Matters
Reduces dependence on expensive labeled datasets, enabling more scalable training of conditional diffusion models.