Research map / Image generation / Fast diffusion sampling
Don't Play Favorites: Minority Guidance for Diffusion Models
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 minority score (Tweedie-based reconstruction discrepancy) and minority guidance, a classifier-guidance sampler that steers diffusion models toward low-density minority samples without labels.
Problem
Diffusion models are majority-oriented: ancestral sampling mostly yields high-likelihood samples, making generation of minority (low-density) samples inefficient, and existing minority samplers need labels.
Approach
Define a minority score as the discrepancy between a clean sample and its Tweedie-denoised reconstruction (one-shot, cheap to compute, equivalent up to scaling to noise-prediction error), then condition the diffusion sampler on categorized minority-score levels via classifier guidance, enabling label-free minority generation with controllable uniqueness.
Key results
- On CelebA 64x64, minority guidance reaches cFID 26.98 and sFID 8.23 versus 75.05 and 16.73 for ADM ancestral sampling and 27.97 cFID for Sehwag et al.
- On LSUN-Bedrooms 256x256, cFID improves to 41.52 from 62.83 for ADM and 56.45 for StyleGAN.
- On the in-house brain MRI, cFID 19.73 and recall 0.47 versus 21.71 and 0.37 for ADM, outperforming StyleGAN2-ADA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use minority guidance instead of the baseline sampler | ADM on CelebA 64×64 · CelebA 64×64 | cFID: 27.97 -> 26.98 | Helped |
| Use minority guidance instead of the baseline sampler | ADM on CelebA 64×64 · CelebA 64×64 | Rec: 0.42 -> 0.34 | Hurt |
| Use minority guidance instead of the baseline sampler | ADM on ImageNet 256×256 · ImageNet 256×256 | sFID: 9.88 -> 7.82 | Helped |
| Use minority guidance instead of the baseline sampler | ADM on ImageNet 256×256 · ImageNet 256×256 | cFID: 13.43 -> 13.83 | Hurt |
| Use minority guidance instead of the baseline sampler | ADM on ImageNet 256×256 · ImageNet 256×256 | Rec: 0.42 -> 0.45 | Helped |
| Use minority guidance instead of the baseline sampler | ADM on LSUN Bedrooms 256×256 · LSUN Bedrooms 256×256 | cFID: 56.45 -> 41.52 | Helped |
| Use minority guidance instead of the baseline sampler | ADM on LSUN Bedrooms 256×256 · LSUN Bedrooms 256×256 | sFID: 5.96 -> 6.67 | Hurt |
| Use minority guidance instead of the baseline sampler | ADM on Brain MRI 256×256 · Brain MRI 256×256 | cFID: 21.71 -> 19.73 | Helped |
| Increase the requested minority class score category | Minority guidance · Generated samples | LOF | Helped |
| Increase classifier scale w while conditioning on minority class l̃ = 0 | Minority guidance · Generated samples | LOF | Hurt |
Limitations
- Requires training a minority classifier, which is computationally expensive for large datasets
- Medical MRI dataset is in-house and not publicly available, limiting reproducibility
- Evaluation of MRI relies on a domain-specific feature space because ImageNet features may not be interpretable there
Experimental setup
Datasets: CelebA CIFAR-10 LSUN-Bedrooms ImageNet CIFAR-10-LT brain MRI (in-house)
Benchmarks: CelebA 64x64 LSUN-Bedrooms 256x256 Brain MRI 256x256 ImageNet 64x64 ImageNet 256x256 CIFAR-10 CIFAR-10-LT
Metrics: AvgkNN LOF Rarity Score cFID sFID Improved Precision Improved Recall
Baselines: BigGAN StyleGAN StyleGAN2-ADA ADM ancestral sampling LDM Sehwag et al.
Builds on: DDPM classifier guidance Tweedie's formula ADM
Closest papers by meaning
- Boost-and-Skip: A Simple Guidance-Free Diffusion for Minority Generation (ICML 2025 · similarity 0.91)
- Provable Efficiency of Guidance in Diffusion Models for General Data Distribution (ICML 2025 · similarity 0.89)
- CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling (ICLR 2024 · similarity 0.89)
- Theoretical insights for diffusion guidance: A case study for Gaussian mixture models (ICML 2024 · similarity 0.89)
- SPARKE: Scalable Prompt-Aware Diversity and Novelty Guidance in Diffusion Models via RKE Score (NeurIPS 2025 · similarity 0.88)
- Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models (NeurIPS 2024 · similarity 0.88)
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.