Research map / Label-efficient and robust learning / Recognition and long-tail
DiffuLT: Diffusion for Long-tail Recognition Without External Knowledge
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.
Trains a diffusion model from scratch only on a long-tailed dataset with a feature-extractor-guided AID loss, generates balanced proxy samples, filters OOD ones, and retrains a classifier with weighted cross-entropy.
Problem
Long-tail image classification suffers from tail-class data scarcity; existing remedies rely on external data or pretrained models, which may be unavailable in practice.
Approach
Identify approximately-in-distribution (AID) generated samples as the most useful for long-tail classification, and bias a diffusion model toward generating them via a feature-space deviation loss using the baseline classifier as feature extractor, then filter OOD samples and train with down-weighted synthetic data.
Key results
- On CIFAR100-LT (r=100), DiffuLT reaches 51.5% accuracy vs. 38.3% CE baseline and 48.9% for H2T; combined with RIDE reaches 52.4%.
- On ImageNet-LT with ResNet-50, DiffuLT reaches 56.4% accuracy vs. 41.6% CE baseline (ResNet-10: 50.4%).
- On CIFAR10-LT (r=100), DiffuLT reaches 84.7% vs. 70.4% CE baseline; ablation shows AID loss lowers FID from 5.86 (CBDM) to 5.37 and raises AID share to 50.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffuLT instead of CE | Long-tail classification · CIFAR100-LT (r=100) | accuracy: 38.3 -> 51.5 | Helped |
| Use DiffuLT instead of CE | Long-tail classification · CIFAR10-LT (r=100) | accuracy: 70.4 -> 84.7 | Helped |
| Use DiffuLT instead of CE | Long-tail classification with ResNet-50 · ImageNet-LT | accuracy: 41.6 -> 56.4 | Helped |
| Add generated samples to training (ablation) | CIFAR100-LT (r=100) classifier · CIFAR100-LT (r=100) | accuracy: 38.3 -> 46.6 | Helped |
| Add the AID-biased loss (ablation) | DiffuLT on CIFAR100-LT (r=100) · CIFAR100-LT (r=100) | accuracy: 46.6 -> 49.7 | Helped |
| Add OOD-sample filtering (ablation) | DiffuLT on CIFAR100-LT (r=100) · CIFAR100-LT (r=100) | accuracy: 49.7 -> 50.3 | Helped |
| Add weighted cross-entropy for generated samples (ablation) | DiffuLT on CIFAR100-LT (r=100) · CIFAR100-LT (r=100) | accuracy: 50.3 -> 51.5 | Helped |
| Increase the generated-sample loss weight to ω=1.0 or higher (ablation) | DiffuLT · CIFAR100-LT (r=100) | accuracy: 51.5; 48.3; 45.1; 43.3 | Hurt |
| Set AID-loss weight α to 0 (ablation) | DiffuLT · CIFAR100-LT (r=100) | accuracy: 51.5 -> 38.3 | Hurt |
| Vary α from the default to 0.5 or 1.0 (ablation) | DiffuLT · CIFAR100-LT (r=100) | accuracy: 51.5 -> 50.3 | Hurt |
| Generate AID samples rather than ID or OOD samples (ablation) | Diffusion-augmented CIFAR100-LT classifier · CIFAR100-LT (r=100) | accuracy: ID 44.2; AID 45.2; OOD 36.2 | Mixed |
| Increase the proportion of head-class data used to train the diffusion model (ablation) | Tail-class generation · CIFAR100-LT tail classes | tail-class accuracy: 26.0; 29.7; 32.5; 32.8 | Helped |
Surprising findings
- Classifiers trained only on AID generated samples gave higher accuracy than those trained on ID or OOD samples, and OOD samples reduced accuracy relative to baseline.
- Lower FID does not always mean better classifier accuracy in isolation; the proportion of AID samples correlated more strongly with accuracy than FID.
Limitations
- Training the diffusion model from scratch is time-consuming (24 h on CIFAR100-LT, ~6 days on ImageNet-LT), limiting scalability to larger datasets.
- Does not surpass methods using large pretrained models such as Stable Diffusion or CLIP, which are excluded by the in-domain setting.
- Code is not yet released; results are limited to three long-tailed benchmarks.
Experimental setup
Datasets: CIFAR10-LT CIFAR100-LT ImageNet-LT
Benchmarks: CIFAR10-LT CIFAR100-LT ImageNet-LT
Metrics: top-1 accuracy FID many/medium/few-shot accuracy
Baselines: Focal Loss LDAM-DRW cRT BBN RIDE CAM-BS MisLAS DiVE CMO SAM CUDA CSA
Builds on: DDPM CBDM BBN DDPM classifier-free guidance-free conditional diffusion
Closest papers by meaning
- LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content (CVPR 2024 · similarity 0.88)
- CORAL: Disentangling Latent Representations in Long-Tailed Diffusion (NeurIPS 2025 · similarity 0.88)
- Harnessing Hierarchical Label Distribution Variations in Test Agnostic Long-tail Recognition (ICML 2024 · similarity 0.88)
- Feature Fusion from Head to Tail for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.88)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.87)
- ADDP: Learning General Representations for Image Recognition and Generation with Alternating Denoising Diffusion Process (ICLR 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.