Research map / Image generation / Fast diffusion sampling
Inversion Circle Interpolation: Diffusion-based Image Augmentation for Data-scarce Classification
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 Diff-II, a diffusion-based data augmentation method that learns per-category concepts, interpolates DDIM inversions on a circle (spherical interpolation/extrapolation), and denoises in two stages with suffixed then plain prompts to jointly improve faithfulness and diversity for data-scarce classification.
Problem
Existing diffusion-based data augmentation for data-scarce image classification struggles to jointly achieve faithfulness (category fidelity) and diversity of synthetic samples.
Approach
Interpolate DDIM inversions of same-category images via random circle (spherical) interpolation so the initial noise stays Gaussian, then use a two-stage denoising with a suffixed prompt (diversity) followed by a plain category-concept prompt (faithfulness).
Key results
- Few-shot: average accuracy gain over original training of up to +10.05% (ResNet-50, 5-shot) and +3.56% (ViT-B/16, 10-shot) across four fine-grained datasets, outperforming all listed diffusion-based DA baselines.
- Long-tail CUB-LT (IF=10): Diff-II reaches 70.28% vs 64.48% for Diff-Mix; average gain +15.74 vs +12.14 over CE.
- OOD Waterbird: average accuracy improved by 11.39% over the original 5-shot CUB training set (46.09 vs 34.70).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Diff-II data augmentation instead of no augmentation | ResNet-50 · 5-shot few-shot classification, average of CUB/Aircraft/Pet/Car | accuracy: 54.56 -> 64.61 | Helped |
| Use Diff-II data augmentation instead of no augmentation | ResNet-50 · 10-shot few-shot classification, average of CUB/Aircraft/Pet/Car | accuracy: 71.11 -> 75.78 | Helped |
| Use Diff-II data augmentation instead of no augmentation | ViT-B/16 · 5-shot few-shot classification, average of CUB/Aircraft/Pet/Car | accuracy: 64.07 -> 72.59 | Helped |
| Use Diff-II data augmentation instead of no augmentation | ViT-B/16 · 10-shot few-shot classification, average of CUB/Aircraft/Pet/Car | accuracy: 77.26 -> 80.82 | Helped |
| Use Diff-II data augmentation instead of the strongest reported baseline, Diff-Mix | ResNet-50 · CUB-LT, average across imbalance factors | accuracy gain over CE: +12.14 -> +15.74 | Helped |
| Use Diff-II data augmentation instead of the strongest reported baseline, Diff-Mix | ResNet-50 · Flower-LT, average across imbalance factors | accuracy gain over CE: +4.53 -> +4.97 | Helped |
| Use Diff-II data augmentation instead of no augmentation | ResNet-50 · Waterbird, average across four land/water groups | accuracy: 34.70 -> 46.09 | Helped |
| Add category concept learning (ablation) | Diff-II on 5-shot Aircraft · Aircraft | accuracy: 31.53 -> 38.74 | Helped |
| Add linear interpolation (ablation) | Concept-learning variant on 5-shot Aircraft · Aircraft | accuracy: 38.74 -> 37.32 | Hurt |
| Add spherical interpolation to linear interpolation (ablation) | Concept-learning variant on 5-shot Aircraft · Aircraft | accuracy: 37.32 -> 38.37 | Helped |
| Add spherical extrapolation to linear interpolation (ablation) | Concept-learning variant on 5-shot Aircraft · Aircraft | accuracy: 37.32 -> 37.70 | Helped |
| Add spherical extrapolation to spherical interpolation (ablation) | Concept-learning variant on 5-shot Aircraft · Aircraft | accuracy: 38.37 -> 40.49 | Helped |
Surprising findings
- Gains shrink as long-tail imbalance grows (IF=100), because categories with a single image cannot be interpolated.
Limitations
- Cannot interpolate for categories with only one training image, reducing gains in extreme long-tail settings.
- Relies on a captioner (BLIP) and an LLM (GPT-4) to extract suffixes, adding pipeline dependency.
- Evaluated only on image classification with domain-specific fine-grained datasets.
Experimental setup
Datasets: CUB Aircraft Pet Cars CUB-LT Flower-LT Waterbird
Benchmarks: CUB Aircraft Pet Cars CUB-LT Flower-LT Waterbird
Metrics: accuracy LPIPS CLIP score
Baselines: Mixup CutMix Real-Filter Real-Guidance Da-Fusion Real-Mix Diff-AUG Diff-Mix CMO CMO-DRW
Builds on: Diff-Mix Da-Fusion DDIM inversion textual inversion LoRA
Closest papers by meaning
- Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model (CVPR 2024 · similarity 0.88)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.87)
- DiffAug: A Diffuse-and-Denoise Augmentation for Training Robust Classifiers (NeurIPS 2024 · similarity 0.87)
- ADDP: Learning General Representations for Image Recognition and Generation with Alternating Denoising Diffusion Process (ICLR 2024 · similarity 0.86)
- DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models (CVPR 2024 · similarity 0.86)
- Few-shot Learner Parameterization by Diffusion Time-steps (CVPR 2024 · similarity 0.85)
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.