[Paper Review] On the Limitation of Diffusion Models for Synthesizing Training Datasets

Shin’ya Yamaguchi, Takuma Fukuda|arXiv (Cornell University)|Nov 22, 2023

Model Reduction and Neural Networks4 citations

TL;DR

This paper investigates why state-of-the-art diffusion models fail to replicate real training datasets for training classifiers, despite generating high-fidelity synthetic samples. By reconstructing real images through controlled reverse diffusion steps, the authors show that synthetic samples increasingly concentrate near data distribution modes, degrading classifier performance due to loss of outlying, high-entropy features.

ABSTRACT

Synthetic samples from diffusion models are promising for leveraging in training discriminative models as replications of real training datasets. However, we found that the synthetic datasets degrade classification performance over real datasets even when using state-of-the-art diffusion models. This means that modern diffusion models do not perfectly represent the data distribution for the purpose of replicating datasets for training discriminative tasks. This paper investigates the gap between synthetic and real samples by analyzing the synthetic samples reconstructed from real samples through the diffusion and reverse process. By varying the time steps starting the reverse process in the reconstruction, we can control the trade-off between the information in the original real data and the information added by diffusion models. Through assessing the reconstructed samples and trained models, we found that the synthetic data are concentrated in modes of the training data distribution as the reverse step increases, and thus, they are difficult to cover the outer edges of the distribution. Our findings imply that modern diffusion models are insufficient to replicate training data distribution perfectly, and there is room for the improvement of generative modeling in the replication of training datasets.

Motivation & Objective

• To investigate the performance degradation when training classifiers solely on synthetic data generated by diffusion models.
• To identify the root cause of the gap between real and synthetic datasets in terms of data distribution coverage.
• To analyze how the reverse denoising process of diffusion models affects sample quality and classifier generalization.
• To evaluate whether synthetic samples preserve the structural and semantic characteristics of real data in feature space.
• To provide insights into improving generative modeling for dataset replication in downstream classification tasks.

Proposed method

• Reconstructing real images by applying a diffusion process up to a pre-defined noise step, then reversing it using the reverse denoising process.
• Controlling the reverse process by varying the starting time step (reverse step) to modulate the trade-off between real and synthetic information.
• Training classifiers on reconstructed synthetic datasets with different reverse steps to evaluate performance degradation.
• Using feature visualization (PCA) and attention maps (GradCAM) to analyze the distribution and behavior of synthetic samples in classifier feature space.
• Measuring output entropy of classifiers to assess sample difficulty and representativeness of synthetic samples.
• Analyzing the relationship between the reverse process and score-based likelihood maximization, showing that each reverse step moves samples toward higher-likelihood (mode) regions.

Experimental results

Research questions

• RQ1 Why do classifiers trained exclusively on synthetic data from diffusion models underperform compared to those trained on real data?
• RQ2 How does the reverse denoising process of diffusion models affect the distribution of synthetic samples in feature space?
• RQ3 To what extent do synthetic samples preserve the structural and semantic diversity of real training data?
• RQ4 Does the reverse process systematically bias samples toward the modes of the data distribution?
• RQ5 Can the reverse process be modified to better cover the outer edges of the data distribution?

Key findings

• Synthetic samples generated by state-of-the-art diffusion models are nearly indistinguishable from real samples in terms of visual quality.
• Classifier performance degrades as the reverse step increases, indicating that more synthetic information leads to worse generalization.
• Synthetic samples are easier to classify than real samples, as shown by lower output entropy in the classifier.
• Feature visualization reveals that synthetic samples concentrate near the modes of the real data distribution and fail to cover the outer edges.
• The reverse denoising process systematically pulls samples toward higher-likelihood regions, explaining the concentration effect.
• The attention maps of classifiers remain consistent across real and synthetic samples, indicating no structural distortion in learned representations.