Research map / Image generation / Fast diffusion sampling
Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DIFFLENS, which uses a k-sparse autoencoder on the U-Net bottleneck activations and gradient-based attribution to find bias features in diffusion models, then scales them to mitigate or control gender, age, and race bias.
Problem
Social biases (gender, age, race) in diffusion models' generated images, where existing mitigation methods work mainly on guidance, prompts, or fine-tuning rather than on the internal mechanisms that causally drive biased outputs.
Approach
Disentangle polysemantic bottleneck neurons into a sparse monosemantic feature space with a k-SAE, attribute the bias classifier's output to each feature via integrated gradients, and intervene (scale or add) only on the top-attributed bias features to control bias level while preserving other attributes.
Key results
- On unconditional P2, DIFFLENS reduces gender FD from 0.226 to 0.002 with FID 31.93 (vs. 33.38 original) and CLIP-I 0.9479.
- On Stable Diffusion, DIFFLENS achieves gender FD 0.046 and age FD 0.049, the best among compared methods, with FID 112.83 and 99.17.
- Ablation on P2 shows bias attribution lowers gender FD from 0.190 (activation) to 0.072, and k-SAE with attribution reaches 0.002.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use k-SAE to disentangle the original latent space into a sparse semantic space (ablation) | DIFFLENS on P2 · P2, gender | CLIP-I: 0.8597 to 0.9479 | Helped |
| use k-SAE to disentangle the original latent space into a sparse semantic space (ablation) | DIFFLENS on P2 · P2, gender | FID: 41.68 to 31.93 | Helped |
| use bias attribution to locate bias features (ablation) | DIFFLENS on P2 · P2, gender | FD: 0.190 to 0.072; 0.303 to 0.002 | Helped |
| mitigate gender bias with DIFFLENS | P2 · P2, gender | FD: 0.226 to 0.002 | Helped |
| mitigate age bias with DIFFLENS | P2 · P2, age | FD: 0.592 to 0.401 | Helped |
| mitigate race bias with DIFFLENS | P2 · P2, race | FD: 0.718 to 0.447 | Helped |
| mitigate gender bias with DIFFLENS | Stable Diffusion · Stable Diffusion, gender | FD: 0.564 to 0.046 | Helped |
| mitigate age bias with DIFFLENS | Stable Diffusion · Stable Diffusion, age | FD: 0.752 to 0.049 | Helped |
| mitigate race bias with DIFFLENS | Stable Diffusion · Stable Diffusion, race | FD: 0.558 to 0.401 | Helped |
Surprising findings
- Latent Editing achieves a lower gender FD (0.003) than DIFFLENS (0.002)? No: DIFFLENS is best at 0.002, but Latent Editing produces artifacts in some images despite strong FD and FID.
- Direct activation-based neuron selection performs poorly for debiasing, indicating polysemanticity hinders neuron-level interventions.
Limitations
- Difficulty identifying certain racial features, as noted in the conclusion.
- Quality declines under intensified edits and at highly skewed gender ratios.
- Evaluated only on gender, age, and race with two diffusion models; the paper suggests extending to Diffusion Transformers.
Experimental setup
Datasets: CelebA-HQ LAION
Benchmarks: CelebA-HQ
Metrics: Fairness Discrepancy (FD) FID CLIP-I CLIP-T
Baselines: Activation Latent Editing H-Distribution Latent Direction Finetuning
Builds on: k-sparse autoencoder integrated gradients Latent Editing H-Distribution
Closest papers by meaning
- Understanding diffusion models requires rethinking (again) generalization (NeurIPS 2026 · similarity 0.83)
- Elucidating the Exposure Bias in Diffusion Models (ICLR 2024 · similarity 0.83)
- Understanding Generalizability of Diffusion Models Requires Rethinking the Hidden Gaussian Structure (NeurIPS 2024 · similarity 0.83)
- An Analytical Theory of Spectral Bias in the Learning Dynamics of Diffusion Models (NeurIPS 2025 · similarity 0.82)
- Training Unbiased Diffusion Models From Biased Dataset (ICLR 2024 · similarity 0.82)
- Generalization through variance: how noise shapes inductive biases in diffusion models (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.