Research map / Image generation / Fast diffusion sampling
Steering Diffusion Models via Class-Contrastive Influence for Few-Shot Classification
Jeeyung Kim, Erfan Esmaeili, Qiang Qiu
Read the paperCheck your idea against 137,217 papers
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 C2I-rewarded RL fine-tuned diffusion augmentation | ViT-B/16 trained on original data · 3 MedMNIST (avg) | AUC: 0.873 -> 0.894 | Improved |
| add C2I-rewarded RL fine-tuning on top of DataDream | DataDream · 3 MedMNIST (avg, ViT) | AUC: 0.866 -> 0.894 | Improved |
| add C2I augmentation to ResNet18 (not used in RL) | ResNet18 · 3 MedMNIST (avg) | AUC: 0.842 -> 0.882 | Improved |
| add C2I augmentation | ViT-B/16 · PneumoniaMNIST | AUC: 0.941 -> 0.945 | No clear effect |
| add C2I augmentation under noisy test inputs | ViT-B/16 · BreastMNIST noisy | AUC: 0.761 -> 0.819 | Improved |
| add C2I augmentation under noisy test inputs | ViT-B/16 · DermaMNIST-binary noisy | AUC: 0.800 -> 0.834 | Improved |
| add C2I augmentation under noisy test inputs | ViT-B/16 · PneumoniaMNIST noisy | AUC: 0.907 -> 0.893 | Hurt |
| add C2I multi-class reward | ViT-B/16 · DermaMNIST-all | accuracy: 0.648 -> 0.683 | Improved |
| lengthen RL fine-tuning from 10 to 20 epochs (ablation) | C2I · BreastMNIST | AUC: 0.83 -> 0.89 | Improved |
| use C2I reward instead of fine-tuning SD on misclassified samples (ablation) | Simple hard-example baseline · BreastMNIST | AUC: 0.778 -> 0.885 | Improved |
| shrink RL validation set from 135 to 16 per class (ablation) | C2I · PneumoniaMNIST | AUC: 0.944 -> 0.943 | No clear effect |
| increase number of synthesized images (ablation) | DataDream and Dataset Expansion · BreastMNIST | AUC | Hurt |
Closest papers by meaning
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.88)
- Are Images Indistinguishable to Humans Also Indistinguishable to Classifiers? (CVPR 2025 · similarity 0.87)
- Diffusion Classifiers Understand Compositionality, but Conditions Apply (NeurIPS 2025 · similarity 0.87)
- Not All Diffusion Model Activations Have Been Evaluated as Discriminative Features (NeurIPS 2024 · similarity 0.87)
- Diffusion-Classifier Synergy: Reward-Aligned Learning via Mutual Boosting Loop for FSCIL (NeurIPS 2025 · similarity 0.86)
- Trade-Offs in Fine-Tuned Diffusion Models between Accuracy and Interpretability (AAAI 2024 · similarity 0.86)
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.
