Research map / Image generation / Fast diffusion sampling
Class-Balancing Diffusion Models
Yiming Qin, Huangjie Zheng, Jiangchao Yao, Mingyuan Zhou, Ya Zhang
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 |
|---|---|---|---|
| Add distribution-adjustment regularization | DDPM · CIFAR100LT | FID: 7.38 -> 6.26 | Improved |
| Add distribution-adjustment regularization | DDPM · CIFAR100 | FID: 3.11 -> 2.72 | Improved |
| Add distribution-adjustment regularization | DDPM · CIFAR10LT | FID: 5.76 -> 5.46 | Improved |
| Add distribution-adjustment regularization | DDPM · CIFAR10 | IS: 9.80 -> 9.63 | Hurt |
| Add ADA augmentation | DDPM · CIFAR100LT | FID: 7.38 -> 6.16 | Improved |
| Add re-sampling or DiffAug instead of ADA | DDPM · CIFAR100LT | FID: 7.38 -> 9.19; 7.38 -> 10.50; 7.38 -> 9.72 | Hurt |
| Add the commitment loss (ablation) | CBDM · CIFAR100LT | FID: 8.84 -> 8.30 | Improved |
| Use a square-root-frequency label set rather than the training-frequency label set (ablation) | CBDM trained from scratch · CIFAR100LT | FID: 6.26 -> 6.00 | Improved |
| Use a balanced label set rather than the training-frequency label set (ablation) | CBDM trained from scratch · CIFAR100LT | Recall: 0.57 -> 0.64 | Improved |
| Use a balanced label set rather than the training-frequency label set (ablation) | CBDM fine-tuned on CIFAR100LT · CIFAR100LT | FID: 6.20 -> 5.85 | Improved |
| Use CBDM-generated images rather than DDPM-generated images for classifier augmentation | ResNet-32 trained on CIFAR100LT plus 50k generated images · CIFAR100 test | Recall: 0.44 -> 0.47 | Improved |
| Use 100-step DDIM sampling rather than DDPM steps (ablation) | CBDM · CIFAR100LT | FID: 6.27 (+0.01) | No clear effect |
Closest papers by meaning
- CORAL: Disentangling Latent Representations in Long-Tailed Diffusion (NeurIPS 2025 · similarity 0.90)
- Balancing Act: Distribution-Guided Debiasing in Diffusion Models (CVPR 2024 · similarity 0.87)
- DiffuLT: Diffusion for Long-tail Recognition Without External Knowledge (NeurIPS 2024 · similarity 0.87)
- Classification Diffusion Models: Revitalizing Density Ratio Estimation (NeurIPS 2024 · similarity 0.87)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.87)
- Slight Corruption in Pre-training Data Makes Better Diffusion Models (NeurIPS 2024 · similarity 0.87)
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.
