Research map / Image generation / Fast diffusion sampling
Generating Risky Samples with Conformity Constraints via Diffusion Models
Han Yu, Hao Zou, Xingxuan Zhang, Zhengyi Wang, Yue He, Kehan Li, Peng Cui
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 embedding screening and gradient guidance together (ablation) | Stable-unCLIP base · NICO++ (ResNet-50) | error rate of generated samples | Improved |
| increase gradient scale s from 0.1 to 50 (ablation) | RiskyDiff · NICO++ (ResNet-50) | error rate of generated samples: 6.9 -> 73.8 | Improved |
| increase gradient scale s beyond 20 (ablation) | RiskyDiff · NICO++ (ResNet-50) | conformity to category: 58.8 -> 73.8 | Hurt |
| increase gradient scale s from 5 to 10 (ablation) | RiskyDiff · NICO++ (ResNet-50) | FID: 17.7 -> 17.5 | Improved |
| raise conformity coefficient from 0 to 0.1 (ablation) | RiskyDiff · NICO++ (ResNet-50) | error rate of generated samples: 41.3 -> 15.9 | Hurt |
| raise conformity coefficient from 0 to 0.0001 (ablation) | RiskyDiff · NICO++ (ResNet-50) | FID: 19.7 -> 17.5 | Improved |
| reduce validation data from full to 0.1 of its size (ablation) | RiskyDiff · NICO++ (ResNet-50) | error rate of generated samples: 41.2 -> 39.4 | No clear effect |
| retrain with AdvDiff-generated data instead of original | ResNet-50 · NICO++ ID accuracy | accuracy: 82.9 -> 80.1 | Hurt |
| retrain with RiskyDiff-generated data | ResNet-50 · PACS OOD accuracy | accuracy: 77.7 -> 81.4 | Improved |
| retrain with Mixup instead of original | ResNet-50 · PACS OOD accuracy | accuracy: 77.7 -> 72.4 | Hurt |
| retrain with RiskyDiff-generated data | ResNet-50 · CIFAR-100 test | accuracy: 78.4 -> 79.7 | Improved |
| retrain with RiskyDiff-generated data | ResNet-50 · NICO++ OOD accuracy | accuracy: 69.6 -> 71.5 | Improved |
Closest papers by meaning
- Slight Corruption in Pre-training Data Makes Better Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Constrained Diffusion Models via Dual Training (NeurIPS 2024 · similarity 0.85)
- Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning (NeurIPS 2025 · similarity 0.84)
- Composition and Alignment of Diffusion Models using Constrained Learning (NeurIPS 2025 · similarity 0.84)
- Constrained Diffusion with Trust Sampling (NeurIPS 2024 · similarity 0.84)
- Zero-Shot Conditioning of Score-Based Diffusion Models by Neuro-Symbolic Constraints (AAAI 2025 · similarity 0.84)
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.
