Training Diffusion-based Generative Models with Limited Data

Zhaoyu Zhang, Yang Hua, Guanxiong Sun, Hui Wang, Seán McLoone

TL;DR: We present a novel theoretical insight for diffusion models and propose a novel diffusion model called Limited Data Diffusion (LD-Diffusion) which can effectively improve the training of diffusion models under limited data settings.

Abstract: Diffusion-based generative models (diffusion models) often require a large amount of data to train a score-based model that learns the score function of the data distribution through denoising score matching. However, collecting and cleaning such data can be expensive, time-consuming, and even infeasible. In this paper, we present a novel theoretical insight for diffusion models that two factors, i.e., the denoiser function hypothesis space and the number of training samples, can affect the denoising score matching error of all training samples. Based on this theoretical insight, it is evident that minimizing the total denoising score matching error is challenging within the denoiser function hypothesis space in existing methods, when training diffusion models with limited data. To address this, we propose a new diffusion model called Limited Data Diffusion (LD-Diffusion), which consists of two main components: a compressing model and a novel mixed augmentation with fixed probability (MAFP) strategy. Specifically, the compressing model can constrain the complexity of the denoiser function hypothesis space and MAFP can effectively increase the training samples by providing more informative guidance than existing data augmentation methods in the compressed hypothesis space. Extensive experiments on several datasets demonstrate that LD-Diffusion can achieve better performance compared to other diffusion models. Codes are available at https://github.com/zzhang05/LD-Diffusion.

Lay Summary: Diffusion models are a type of AI model that can generate realistic images, sounds, or other data by learning from large datasets. However, training these models usually requires a lot of clean and high-quality data, which can be difficult or expensive to collect. In this paper, we explore why diffusion models struggle with limited data: the model’s ability to learn is limited both by the amount of training data and the complexity of the denoiser function hypothesis space during training. Based on this insight, we introduce a new method called LD-Diffusion. Our approach includes two innovations: a compressing model that constrains the complexity of the denoiser function hypothesis space, and a new data augmentation technique that helps the model see more useful variations of the data. Our experiments show that LD-Diffusion performs better than existing diffusion models when training data is scarce. The code for our method is available online.

Primary Area: Deep Learning->Generative Models and Autoencoders

Keywords: Diffusion models; Limited Data; LD-Diffusion

Submission Number: 8061