# Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability

arXiv: https://arxiv.org/abs/2503.20483 · HTML: https://mlsanity.com/paper/2503.20483/

## Key facts

- Cited by 26 papers in the map (data through 14 Oct 2026).
- Placed in Image generation, text-to-image and editing, one of 14 directions in Diffusion models.
- First posted on arXiv 26 Mar 2025.

## Placed in

- Diffusion models / [Image generation, text-to-image and editing](https://mlsanity.com/diffusion/image-generation-text-to-image-and-editing/)
- Mechanistic interpretability / [Probing deception, bias and beliefs](https://mlsanity.com/mechinterp/probing-deception-bias-and-beliefs/)

## Built on by (most cited first, 26 of 26)

- [Temporal Concept Dynamics in Diffusion Models via Prompt-Conditioned Interventions](https://mlsanity.com/paper/2512.08486/) (2025)
- [DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders](https://mlsanity.com/paper/2602.05859/) (2026)
- [Let EEG Models Learn EEG](https://mlsanity.com/paper/2605.21280/) (2026)
- [Interpretable Debiasing of Vision-Language Models for Social Fairness](https://mlsanity.com/paper/2602.24014/) (2026)
- [Revealing the Attention Floating Mechanism in Masked Diffusion Models](https://mlsanity.com/paper/2601.07894/) (2026)
- [How I Met Your Bias: Investigating Bias Amplification in Diffusion Models](https://mlsanity.com/paper/2512.20233/) (2025)
- [Fully Unsupervised Self-debiasing of Text-to-Image Diffusion Models](https://mlsanity.com/paper/2512.03749/) (2025)
- [BiasMap: Leveraging Cross-Attentions to Discover and Mitigate Hidden Social Biases in Text-to-Image Generation](https://mlsanity.com/paper/2509.13496/) (2025)
- [FAME: Fairness-aware Attention-modulated Video Editing](https://mlsanity.com/paper/2510.22960/) (2025)
- [SCALEX: Scalable Concept and Latent Exploration for Diffusion Models](https://mlsanity.com/paper/2511.13750/) (2025)
- [Locating Demographic Bias at the Attention-Head Level in CLIP's Vision Encoder](https://mlsanity.com/paper/2603.11793/) (2026)
- [Attention, May I Have Your Decision? Localizing Generative Choices in Diffusion Models](https://mlsanity.com/paper/2604.06052/) (2026)
- [How Bias Binds: Measuring Hidden Associations for Bias Control in Text-to-Image Compositions](https://mlsanity.com/paper/2511.07091/) (2025)
- [Aligned but Stereotypical? How System Prompts Shape Demographic Bias in LLM-Based Text-to-Image Models](https://mlsanity.com/paper/2512.04981/) (2025)
- [FAIRT2V: Training-Free Debiasing for Text-to-Video Diffusion Models](https://mlsanity.com/paper/2601.20791/) (2026)
- [Exposing Hidden Biases in Text-to-Image Models via Automated Prompt Search](https://mlsanity.com/paper/2512.08724/) (2025)
- [RAIGen: Rare Attribute Identification in Text-to-Image Generative Models](https://mlsanity.com/paper/2602.06806/) (2026)
- [The Latent Color Subspace: Emergent Order in High-Dimensional Chaos](https://mlsanity.com/paper/2603.12261/) (2026)
- [Attention Frequency Modulation: Training-Free Spectral Modulation of Diffusion Cross-Attention](https://mlsanity.com/paper/2603.28114/) (2026)
- [Benchmarking Fairness in Spiking Neural Networks: Data Bias, Spurious Features, and Hardware Effects](https://mlsanity.com/paper/2605.27407/) (2026)
- [FairFlow: Demystifying and Mitigating Stereotype Bias in Text-to-Image Diffusion Transformers](https://mlsanity.com/paper/2607.03180/) (2026)
- [Efficient bias mitigation in T2I diffusion models using Concept Graphs](https://mlsanity.com/paper/2607.03397/) (2026)
- [AEGIS: A Mechanism-Guided Defense against Visual Synonym Jailbreaks in Text-to-Image Models](https://mlsanity.com/paper/2607.06120/) (2026)
- [Introspective Attention Modulation for Safe Text-to-Image Generation](https://mlsanity.com/paper/2607.14945/) (2026)
- [ContextBias: Controlled Evaluation of Bias Persistence Under Context Shift in Text-to-Image Models](https://mlsanity.com/paper/2608.29847/) (2026)
- [Debias Anything: Fairness with Diversity without Supervision in Diffusion Models](https://mlsanity.com/paper/2610.01815/) (2026)

## Abstract (arXiv)

Diffusion models have demonstrated impressive capabilities in synthesizing diverse content. However, despite their high-quality outputs, these models often perpetuate social biases, including those related to gender and race. These biases can potentially contribute to harmful real-world consequences, reinforcing stereotypes and exacerbating inequalities in various social contexts. While existing research on diffusion bias mitigation has predominantly focused on guiding content generation, it often neglects the intrinsic mechanisms within diffusion models that causally drive biased outputs. In this paper, we investigate the internal processes of diffusion models, identifying specific decision-making mechanisms, termed bias features, embedded within the model architecture. By directly manipulating these features, our method precisely isolates and adjusts the elements responsible for bias generation, permitting granular control over the bias levels in the generated content. Through experiments on both unconditional and conditional diffusion models across various social bias attributes, we demonstrate our method's efficacy in managing generation distribution while preserving image quality. We also dissect the discovered model mechanism, revealing different intrinsic features controlling fine-grained aspects of generation, boosting further research on mechanistic interpretability of diffusion models.