Research map / Image generation / Fast diffusion sampling
Fair Sampling in Diffusion Models through Switching Mechanism
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 attribute switching, a training-free diffusion sampling method that switches the sensitive-attribute condition at a transition point chosen so that the generated data becomes independent of the attribute (epsilon-fairness).
Problem
Diffusion models amplify training-data bias in sensitive attributes, and existing fairness measures for generative models rely on downstream classifiers or label-based notions, so a classifier-free measure of data fairness for diffusion sampling is needed.
Approach
During reverse diffusion, the sensitive-attribute condition is switched from s0 to s1 at a transition point tau chosen by equalizing cumulative score differences between the two conditions; high-level features from s0 are kept while the attribute is changed, making the output distribution independent of S without extra training.
Key results
- On FairFace train-to-test error rates, attribute switching reaches BER 50.73% (Orig Tr to Syn Te) and 54.78% (Syn Tr to Orig Te), versus 15.14% and 8.50% for vanilla sampling.
- Attribute switching keeps total FID close to vanilla sampling (e.g., FairFace total 31.11 vs 29.50 vanilla) while reducing the S-group FID gap.
- On CelebA with a Smiling classifier, switching keeps smiling accuracy at 88.14% (vanilla 88.38%) while Male accuracy drops to 55.71% (vanilla 93.04%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use attribute switching instead of vanilla sampling | Diffusion sampling · FairFace Race, Orig (Tr) → Syn (Te) | BER (%): 15.14 -> 50.73 | Helped |
| use attribute switching instead of vanilla sampling | Diffusion sampling · FairFace Race, Syn (Tr) → Orig (Te) | BER (%): 8.50 -> 54.78 | Helped |
| use attribute switching instead of mixing conditional embeddings | Diffusion sampling · FairFace Race, Orig (Tr) → Syn (Te) | BER (%): 41.60 -> 50.73 | Helped |
| use attribute switching instead of DiffEdit | Diffusion sampling · FairFace Race, Orig (Tr) → Syn (Te) | BER (%): 21.24 -> 50.73 | Helped |
| use attribute switching instead of vanilla sampling | Diffusion sampling · FairFace Race (12k), total | FID: 29.50 -> 31.11 | No clear effect |
| use attribute switching instead of vanilla sampling | Diffusion sampling · CelebA Male (12k), total | FID: 7.19 -> 7.83 | No clear effect |
| use attribute switching instead of vanilla sampling | Diffusion sampling · CelebA Heavy-Makeup (12k), total | FID: 7.19 -> 7.59 | No clear effect |
| use attribute switching instead of vanilla sampling | CelebA synthetic-data classifier · Smiling task | accuracy (%): 88.38 -> 88.14 | No clear effect |
| use attribute switching instead of vanilla sampling | CelebA synthetic-data classifier · Male attribute task | accuracy (%): 93.04 -> 55.71 | Mixed |
Surprising findings
- Vanilla sampling's BER on the Syn-to-Orig evaluation is 8.50%, lower than the 'Real' baseline BER, indicating the classifier-based metric is sensitive to evaluation direction.
Limitations
- Elements with significant contextual impact can be removed during switching.
- Focuses on distributional fairness for binary sensitive attributes, one facet of fairness.
- Transition point is chosen from a single mini-batch and assumed to transfer to the full sampling process.
Experimental setup
Datasets: FairFace CelebA CIFAR-10
Benchmarks: FairFace CelebA CIFAR-10
Metrics: balanced error rate (BER) classifier accuracy gap Frechet Inception Distance (FID) PCA visualization
Baselines: Vanilla sampling Mixing conditional embeddings DiffEdit
Builds on: DDPM DDIM epsilon-fairness DiffEdit
Closest papers by meaning
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.87)
- Elucidating the Exposure Bias in Diffusion Models (ICLR 2024 · similarity 0.87)
- CoInD: Enabling Logical Compositions in Diffusion Models (ICLR 2025 · similarity 0.87)
- The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.86)
- Balancing Act: Distribution-Guided Debiasing in Diffusion Models (CVPR 2024 · similarity 0.86)
- Bias Mitigation in Graph Diffusion Models (ICLR 2025 · similarity 0.86)
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.