Research map / Image generation / Text-to-image diffusion
Harnessing Diffusion-Generated Synthetic Images for Fair Image Classification
Abhipsa Basu, Aviral Gupta, Abhijnya Bhat, Venkatesh Babu Radhakrishnan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use last-layer retraining on real data after vanilla SD pretraining instead of FFR full finetuning | Vanilla SD pipeline · Waterbirds, CelebA, UTKFace | accuracy: outperforms FFR by an average of 2.4% | Improved |
| cluster each group and train a Dreambooth model per cluster | Dreambooth · UTKFace (original) | worst-group accuracy: 57.9 -> 76.0 | Improved |
| use Clustered Dreambooth instead of LoRA finetuning | LoRA-finetuning · UTKFace (original) | worst-group accuracy: improving upon LoRA by 7.4% | Improved |
| cluster groups into multiple Dreambooth models (averaged over datasets and ratios) | Dreambooth · three fairness benchmarks, both bias ratios | worst-group accuracy: 76.2 -> 79.1 | Improved |
| pretrain on Clustered Dreambooth images before GDRO finetuning | GDRO · Waterbirds (bias ratio 0.999) | worst-group accuracy: 23.5 -> 81.8 | Improved |
| pretrain on Clustered Dreambooth images before GDRO finetuning | GDRO · CelebA (bias ratio 0.999) | worst-group accuracy: 27.2 -> 64.4 | Improved |
| use finetuned generators (LoRA, Dreambooth variants) instead of vanilla SD for bias-conflicting UTKFace images | Clustered Dreambooth · UTKFace (bias ratio 0.999) | worst-group accuracy: 67.8 -> 60.5 | Hurt |
| add CLIP-Centroid to the CLIP-Label filtering score (alpha 1 -> 0.5) (ablation) | Clustered Dreambooth · UTKFace (original) | worst-group accuracy: 72.7 -> 76.0 | Improved |
| use CLIP-Score filtering instead of random image selection (ablation) | Clustered Dreambooth · Waterbirds (original) | worst-group accuracy: 84.9 -> 88.1 | Improved |
| add the real-data finetuning stage after synthetic pretraining instead of mixing real and synthetic data in one stage (ablation) | Clustered Dreambooth · CelebA (original) | worst-group accuracy: 48.3 -> 84.1 | Improved |
| add SupCon loss (beta 1 -> 0.5 in both stages) (ablation) | Clustered Dreambooth · UTKFace (bias ratio 0.999) | worst-group accuracy: 49.6 -> 60.5 | Improved |
| add SupCon loss (beta 1 -> 0.5 in both stages) (ablation) | Clustered Dreambooth · Waterbirds (original) | worst-group accuracy: 88.2 -> 88.1 | No clear effect |
Closest papers by meaning
- Balancing Act: Distribution-Guided Debiasing in Diffusion Models (CVPR 2024 · similarity 0.89)
- Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision (AAAI 2024 · similarity 0.88)
- Finetuning Text-to-Image Diffusion Models for Fairness (ICLR 2024 · similarity 0.87)
- CORAL: Disentangling Latent Representations in Long-Tailed Diffusion (NeurIPS 2025 · similarity 0.87)
- Stay Fair! Ensuring Group Fairness in Diffusion Models Across Guidance Scales (NeurIPS 2026 · similarity 0.87)
- Slight Corruption in Pre-training Data Makes Better Diffusion 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.
