Research map / Image generation / Fast diffusion sampling
Distribution-Aware Data Expansion with 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 DistDiff, a training-free diffusion-based data expansion method that uses hierarchical (class- and group-level) prototypes to steer intermediate denoising steps toward the real data distribution.
Problem
Synthetic data from diffusion models for dataset expansion can deviate from the real data distribution, hurting downstream classifier performance on small datasets; prior fixes need fine-tuning or only filter outputs.
Approach
Build class-level and group-level (agglomerative clustering) feature prototypes from the original data, then apply residual multiplicative latent transforms and optimize intermediate predicted clean latents with hierarchical energy guidance during Stable Diffusion sampling, with no model training.
Key results
- On Caltech-101 with 5x expansion, DistDiff reaches 83.09% accuracy vs 66.71% for the original data and 76.57% for plain SD expansion (ResNet-50 classifier).
- DistDiff outperforms GIF-SD by 5.46% accuracy on Caltech-101 and all transformation baselines in the Caltech-101 comparison (e.g., 83.38 with default augmentation).
- With stronger pretrained baselines, DistDiff improves StanfordCars by 3.56% and PathMNIST by 5.18% over the original; ImageNette and Caltech-101 gains are marginal.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| expand with DistDiff instead of training on original data | ImageNette with ImageNet-1k pretrained ResNet-50 · ImageNette | accuracy: 99.40 -> 99.30 | No clear effect |
| expand with DistDiff instead of training on original data | Caltech-101 with ImageNet-1k pretrained ResNet-50 · Caltech-101 | accuracy: 96.87 -> 97.00 | Helped |
| expand with DistDiff instead of training on original data | StanfordCars with ImageNet-1k pretrained ResNet-50 · StanfordCars | accuracy: 87.61 -> 91.17 | Helped |
| expand with DistDiff instead of training on original data | PathMNIST with CLIP-ViT-B/32 · PathMNIST | accuracy: 84.29 -> 89.47 | Helped |
| add class-level prototype guidance (ablation) | DistDiff with group-level prototype guidance · Caltech-101 | accuracy: $82.70\pm0.07$ -> $83.09\pm0.11$ | Helped |
| add group-level prototype guidance (ablation) | DistDiff with class-level prototype guidance · Caltech-101 | accuracy: $82.84\pm0.54$ -> $83.09\pm0.11$ | Helped |
| remove both hierarchical prototype guidance levels (ablation) | DistDiff · Caltech-101 | accuracy: $83.09\pm0.11$ -> $76.57\pm0.35$ | Hurt |
| vary optimization step M (ablation) | DistDiff · Caltech-101 | accuracy: 1: $81.54 \pm 0.32$; 10: $82.21 \pm 0.22$; 20: $82.36 \pm 0.05$; 25: $82.11 \pm 0.55$ | Mixed |
| vary group-level prototype count K (ablation) | DistDiff · Caltech-101 | accuracy: 2: $82.69 \pm 0.51$; 3: $83.09 \pm 0.11$; 4: $83.08 \pm 0.13$; 5: $83.08 \pm 0.21$ | Mixed |
| increase semantic-stage optimization steps (ablation) | DistDiff · Caltech-101 | accuracy: 1: $82.54 \pm 0.54$; 2: $82.94 \pm 0.43$; 3: $82.77 \pm 0.19$; 4: $82.55 \pm 0.30$ | Mixed |
| replace residual multiplicative transformation with direct latent-point guidance (ablation) | DistDiff · Caltech-101 | accuracy: 76.77% vs 83.19% | Hurt |
| vary guidance learning rate ρ (ablation) | DistDiff · Caltech-101 | accuracy: 0.1: $82.49 \pm 0.33$; 1.0: $82.74 \pm 0.32$; 10.0: $83.09 \pm 0.11$; 20.0: $82.46 \pm 0.35$ | Mixed |
Surprising findings
- On ImageNette and Caltech-101 with ImageNet-pretrained models, SD expansion slightly decreases or barely changes accuracy, and the expansion gives little benefit when the baseline is already strong.
- Directly guiding latent points without the residual multiplicative transform reached only 76.77%, barely above plain Stable Diffusion.
Limitations
- Adds extra computation time that grows as data scales up.
- Requires an auxiliary guidance model (ResNet-50 trained from scratch) that adds development cost.
- Gains are marginal when the baseline model is strong (e.g., ImageNette, Caltech-101 with ImageNet pretraining).
Experimental setup
Datasets: Caltech-101 CIFAR100-Subset ImageNette StandardCars DTD PathMNIST ImageNet-1k LAION
Benchmarks: Caltech-101 CIFAR100-Subset ImageNette StandardCars DTD PathMNIST
Metrics: top-1 accuracy FID-3K
Baselines: AutoAugment RandAugment Random Erasing GridMask MixUp CutMix Stable Diffusion 1.4 LECF GIF-SD
Builds on: Stable Diffusion GIF DDIM CLIP
Closest papers by meaning
- Taming Diffusion for Dataset Distillation with High Representativeness (ICML 2025 · similarity 0.89)
- MGD$^3$ : Mode-Guided Dataset Distillation using Diffusion Models (ICML 2025 · similarity 0.89)
- D^4: Dataset Distillation via Disentangled Diffusion Model (CVPR 2024 · similarity 0.88)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Efficient Dataset Distillation via Minimax Diffusion (CVPR 2024 · similarity 0.88)
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · 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.