Research map / Image generation / Fast diffusion sampling
Multi-scale Diffusion Denoised Smoothing
Jongheon Jeong, Jinwoo Shin
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 cascaded smoothing over sigma 0.25-1.00 | diffusion denoised smoothing · CIFAR-10 sigma=1.0 | certified accuracy at eps=0.0: 48.6 -> 79.6 | Improved |
| add cascaded smoothing over sigma 0.25-1.00 | diffusion denoised smoothing · ImageNet sigma=1.0 | ACR: 0.538 -> 0.752 | Improved |
| add diffusion calibration fine-tuning | diffusion denoised smoothing · CIFAR-10 sigma=0.5 | ACR: 0.646 -> 0.664 | Improved |
| add diffusion calibration fine-tuning | diffusion denoised smoothing · CIFAR-10 sigma=1.0 | empirical accuracy C10: 87.3 -> 83.3 | Hurt |
| add calibration on top of cascading | cascaded smoothing · ImageNet 0.25-1.00 | ACR: 0.752 -> 0.775 | Improved |
| replace cross-entropy with Brier loss for diffusion fine-tuning (ablation) | cascaded smoothing · CIFAR-10 | ACR: 0.605 -> 0.666 | Improved |
| add anti-consistency loss to Brier loss (alpha=1.0) (ablation) | cascaded smoothing · CIFAR-10 | certified accuracy at eps=0.00: 77.6 -> 79.6 | Mixed |
| fine-tune diffusion model only (ablation) | cascaded smoothing · CIFAR-10 sigma=0.5 | ACR: 0.639 -> 0.664 | Improved |
| fine-tune both diffusion model and classifier (ablation) | cascaded smoothing · CIFAR-10 sigma=0.5 | ACR: 0.639 -> 0.673 | Improved |
| increase lambda from 0.001 to 0.1 (alpha=1.0) (ablation) | anti-consistency regularization · CIFAR-10 | empirical clean accuracy: 84.1 -> 80.8 | Hurt |
| increase alpha from 1.0 to 8.0 (lambda=0.001) (ablation) | anti-consistency regularization · CIFAR-10 | certified accuracy at eps=0.00: 79.9 -> 81.5 | Improved |
| raise abstention threshold p0 from 0.5 to 0.6 (ablation) | cascaded smoothing · CIFAR-10 | certified accuracy at eps=0.0: 81.1 -> 83.9 | Improved |
Closest papers by meaning
- Robust Representation Consistency Model via Contrastive Denoising (ICLR 2025 · similarity 0.90)
- Your Diffusion Model is Secretly a Noise Classifier and Benefits from Contrastive Training (NeurIPS 2024 · similarity 0.87)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.86)
- Interpreting and Improving Diffusion Models from an Optimization Perspective (ICML 2024 · similarity 0.86)
- Robust Image Denoising through Adversarial Frequency Mixup (CVPR 2024 · similarity 0.85)
- Training-Free Safe Denoisers for Safe Use of Diffusion Models (NeurIPS 2025 · 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.
