Research map / Image generation / Autoregressive and latent generation
Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing
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.
Diffusing DeBias trains a class-conditional diffusion model on biased data, samples synthetic bias-aligned images, and uses a bias amplifier trained only on them as a plug-in for two unsupervised debiasing recipes (two-step G-DRO and end-to-end LfF-style).
Problem
Unsupervised debiasing of image classifiers trained on data with strong spurious correlations; auxiliary bias models overfit and memorize bias-conflicting samples, degrading the bias signal.
Approach
Train a class-conditional diffusion model (with classifier-free guidance) on the biased dataset, which amplifies per-class bias; sample synthetic bias-aligned images and train the bias amplifier only on them so it never sees real bias-conflicting samples, then plug it into G-DRO (pseudo-labels) or LfF-style reweighting.
Key results
- DDB-II reaches 91.56% WGA on Waterbirds, slightly above supervised G-DRO (91.40%) and above all unsupervised baselines (best prior 90.70%).
- DDB-I reaches 74.67% conflicting accuracy on BFFHQ, versus 73.60% for ETF-Debias.
- DDB-II reaches 72.81% average accuracy on BAR, versus 69.88% for DebiAN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DDB-II instead of the strongest listed unsupervised baseline | Waterbirds debiasing · Waterbirds | WGA: 89.43 -> 91.56 | Helped |
| Use DDB-II instead of the strongest listed baseline | BAR debiasing · BAR | Avg. accuracy: 69.88 -> 72.81 | Helped |
| Use DDB-I instead of ETF-Debias | BFFHQ debiasing · BFFHQ | Conflicting accuracy: 73.60 -> 74.67 | Helped |
| Use DDB-I instead of LWBC | ImageNet-A debiasing · ImageNet-A | Avg. accuracy: 35.97 -> 39.80 | Helped |
| Use DDB-I instead of LLE | UrbanCars debiasing · UrbanCars | BG-Gap: 2.10 -> 1.85 | Helped |
| Increase synthetic images used to train the Bias Amplifier from 100 to 2000 per class (ablation) | DDB-I · Waterbirds | WGA: 86.25 -> 90.81 | Helped |
| Set CFG guidance strength to w=1 rather than w=0 (ablation) | DDB-I · Waterbirds | WGA: 87.85 -> 90.81 | Helped |
| Use plain error-set selection instead of filtered bias-conflicting samples (ablation) | DDB-I · Waterbirds | WGA: 90.81 -> 90.34 | Hurt |
| Increase synthetic images used to train the Bias Amplifier from 100 to 2000 total (ablation) | DDB-II · Waterbirds | WGA: 84.42 -> 91.56 | Helped |
| Set CFG guidance strength to w=1 rather than w=0 (ablation) | DDB-II · Waterbirds | WGA: 87.07 -> 91.56 | Helped |
| Apply DDB-I rather than ERM | Waterbirds with varying bias correlation · Waterbirds | WGA: 62.60 -> 90.81; 63.40 -> 86.91; 64.12 -> 90.19; 68.84 -> 86.14 | Helped |
| Train the Bias Amplifier on real training data instead of synthetic data (ablation) | DDB-I, Recipe I · Waterbirds | WGA: 90.81 -> 79.43 | Hurt |
Surprising findings
- Training the bias amplifier on real data memorized all bias-conflicting samples within 50 epochs in Recipe I, leaving empty pseudo-label groups.
- The diffusion model amplifies bias: synthetic data on Waterbirds reached rho=0.99 when trained on rho=0.95 data.
Limitations
- High computational cost of training diffusion models (about 14 hours on an NVIDIA A30 for BFFHQ at 64x64).
- Generation quality may degrade when the training dataset is very small.
- Lack of large-scale bias-annotated datasets for evaluation.
Experimental setup
Datasets: Waterbirds BFFHQ BAR ImageNet-9 UrbanCars CIFAR-10
Benchmarks: Waterbirds BAR BFFHQ ImageNet-A UrbanCars CIFAR-10
Metrics: worst group accuracy average accuracy conflicting accuracy ID.ACC and accuracy gaps (UrbanCars)
Baselines: ERM G-DRO LISA JTT LfF DebiAN ETF-Debias MoDAD
Builds on: Learning from Failure (LfF) Group DRO Just Train Twice (JTT) Classifier-free guidance
Closest papers by meaning
- Balancing Act: Distribution-Guided Debiasing in Diffusion Models (CVPR 2024 · similarity 0.88)
- A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective (NeurIPS 2024 · similarity 0.87)
- An Effective Theory of Bias Amplification (ICLR 2025 · similarity 0.87)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.86)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.86)
- Debiasing Synthetic Data Generated by Deep Generative Models (NeurIPS 2024 · 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.