How I Met Your Bias: Investigating Bias Amplification in Diffusion Models

TL;DR

This work demonstrates that bias amplification in diffusion-generated images is not solely determined by training data or model architecture, but can be significantly influenced by sampling hyperparameters. By formalizing bias and proposing a training-free bias estimator, the authors show that conditioning strength, integration steps, and stochasticity can push outputs toward bias-aligned or bias-conflicting representations across multiple datasets and diffusion frameworks. The key finding is that bias can both increase and decrease with different samplers and settings, suggesting practical debiasing opportunities through sampling strategies without retraining. These insights offer a complementary avenue to debiasing methods and highlight the need for careful baselines and theoretical understanding of the sampling dynamics in diffusion models.

Abstract

Diffusion-based generative models demonstrate state-of-the-art performance across various image synthesis tasks, yet their tendency to replicate and amplify dataset biases remains poorly understood. Although previous research has viewed bias amplification as an inherent characteristic of diffusion models, this work provides the first analysis of how sampling algorithms and their hyperparameters influence bias amplification. We empirically demonstrate that samplers for diffusion models -- commonly optimized for sample quality and speed -- have a significant and measurable effect on bias amplification. Through controlled studies with models trained on Biased MNIST, Multi-Color MNIST and BFFHQ, and with Stable Diffusion, we show that sampling hyperparameters can induce both bias reduction and amplification, even when the trained model is fixed. Source code is available at https://github.com/How-I-met-your-bias/how_i_met_your_bias.

Figures (23)

• Figure 1: In this work, we investigate the impact of the sampling parameters in (trained) diffusion models. Without retraining, some parameters impact the generation of biased data.
• Figure 2: The integration scheme in edm. In the window $[S_\text{tmin}, S_\text{tmax}]$ we first add independent noise to the image before performing denoising. Outside of the window (or when $~{S_\text{churn}=0}$) we simply perform a deterministic denoising step. The denoising step is represented with a plain arrow, while the step of adding noise is represented with a dotted arrow. The denoising step can either be performed using Euler's method (calling once the network) or Heun's method (calling twice the network).
• Figure 3: CFG on BFFHQ ($\rho_\text{dataset}\mathord{=}0.995$), standard DDPM sampler, varying the number of T.
• Figure 4: $\rho_\text{model}$ vs guidance scale of CFG on 2-classes Biased MNIST ($\rho_\text{dataset}\mathord{=}0.99$) using Karras custom stochastic sampler.
• Figure 5: $\rho_\text{model}$ vs guidance scale of CFG for various T on 2-classes Biased MNIST ($\rho_\text{dataset}\mathord{=}0.90$), standard DDPM sampler.