Deeper Diffusion Models Amplify Bias

Shahin Hakemi, Naveed Akhtar, Ghulam Mubashar Hassan, Ajmal MianarXiv 2025 · DOI · arXiv

Abstract

Generative Diffusion Models (DMs) exhibit a fundamental bias-variance tradeoff dictated by model depth. While deeper models produce more diverse and realistic images, they significantly amplify inherent biases in training data by abstracting away minority features. Conversely, shallower models remain more faithful to the original data distribution but risk compromising privacy through the over-memorization of training samples.

Introduction

Generative Diffusion Models have revolutionized AI-driven imagery, yet their internal mechanics often operate as a "black box." Two primary concerns have emerged: social bias (e.g., over-representing specific demographics) and privacy (the "memorization" of exact training images). This research addresses these issues by framing DM behavior through the lens of the classic bias-variance tradeoff, exploring how model depth—and the resulting level of abstraction—influences whether a model generalizes fairly or simply repeats its training data.

Diffusion Models as Implicit Memory

The authors propose a novel conceptual framework: viewing Diffusion Models as Implicit Memory Modules. This is best understood through a $k$-Nearest Neighbors ($k$-NN) analogy:

  • Low Abstraction (Small $k$): Similar to a shallow model, the system focuses on specific, individual data points. This results in high variance and a high likelihood of memorization.
  • High Abstraction (Large $k$): Similar to a deep model, the system "smooths" the data landscape, attending to many neighbors simultaneously. This leads to high bias and broader generalization.

Mathematically, the DM's latent space is modeled as an "energy landscape" where training points are local minima (valleys). The researchers prove that higher abstraction (deeper models) creates smoother landscapes. However, when these landscapes become too smooth, nearby valleys representing minority classes merge into the larger valleys of the majority class, predisposed the model to generate majority-class outputs.

Visual evidence of bias amplification: as model depth increases, minority class features (e.g., blond hair on males) are smoothed into majority class features.

Experimental Results

The study utilized unconditional DMs with varying depths (2 to 6 encoding blocks) tested against biased versions of the MNIST and CelebA datasets.

Bias Amplification in CelebA

In the CelebA dataset, where the training data was 58.32% female, a shallow 2-block model stayed relatively close to the source distribution (59.67%). However, a deep 6-block model amplified this majority to 71.15%. This effect was even more pronounced for extreme minority groups, such as "Blond Males" (<1%), which were frequently transformed into "Non-blond Males" or "Females" as depth increased.

The Diversity-Fidelity Gap

Quantitative metrics confirm this shift. Deeper models achieved higher LPIPS scores (indicating greater image diversity) but lower Cosine Similarity to the training set. This suggests that while deeper models are more "creative," they are less "honest" about the actual data distribution.

Quantitative relationship between model depth, minority class representation, and image diversity (LPIPS).

The Privacy-Fairness Tug-of-War

A critical insight of this work is the literal tug-of-war between privacy and fairness. To protect privacy by reducing memorization, one might increase model abstraction; however, this inadvertently amplifies social bias by erasing minority groups from the output.

Conversely, the authors found they could "force" a pre-trained model (like Stable Diffusion v1.5) into a high-variance state by bypassing its mid-blocks. This caused the model to generate images that more closely resembled specific training patterns, highlighting the inherent risk of data leakage in shallower architectures.

Conclusion

Striving for "better generalization" in AI is not without cost. This research warns that excessive abstraction leads to a "High Bias" state where the model effectively averages out diversity. Finding the "Goldilocks" zone—the perfect depth that balances privacy and fairness—remains a vital challenge for future architectural research in generative modeling.

Summary updated

Abstract

Despite the remarkable performance of generative Diffusion Models (DMs), their internal working is still not well understood, which is potentially problematic. This paper focuses on exploring the important notion of bias-variance tradeoff in diffusion models. Providing a systematic foundation for this exploration, it establishes that at one extreme, the diffusion models may amplify the inherent bias in the training data, and on the other, they may compromise the presumed privacy of the training samples. Our exploration aligns with the memorization-generalization understanding of the generative models, but it also expands further along this spectrum beyond "generalization", revealing the risk of bias amplification in deeper models. Our claims are validated both theoretically and empirically.
BibTeX
@misc{hakemi2025deeper,
  author = {Shahin Hakemi and Naveed Akhtar and Ghulam Mubashar Hassan and Ajmal Mian},
  title = {{Deeper Diffusion Models Amplify Bias}},
  year = {2025},
  eprint = {2505.17560},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.17560},
}
Figure