Research map / Image generation / Fast diffusion sampling
Moderating the Generalization of Score-based Generative Model
Wan Jiang, He Wang, Xin Zhang, Dan Guo, Zhaoxin Fan, Yunfeng Diao, Richang Hong
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 |
|---|---|---|---|
| Apply MSGM score adjustment instead of re-training only on retained data | VE SDE · Toy Gaussian mixture, NSFG data D_f | NLL on D_f: 11.59 -> 39.01 | Improved |
| Use Orthogonal-MSGM instead of Unseen by Re-training | VE SDE · MNIST classes 3 and 7 | Unlearning ratio (%): 4.1 -> 1.2 | Improved |
| Use Orthogonal-MSGM instead of Unseen by Re-training | VE SDE · MNIST classes 3 and 7 | NLL on D_f: 3.01 -> 13.23 | Improved |
| Use Obtuse-MSGM instead of Unseen by Re-training | VP SDE · CelebA bangs | Unlearning ratio (%): 6.7 -> 0.7 | Improved |
| Use Obtuse-MSGM instead of Orthogonal-MSGM for feature unlearning | VP SDE · CelebA bangs | Unlearning ratio (%): 3.5 -> 0.7 | Improved |
| Use Orthogonal-MSGM instead of Obtuse-MSGM for class unlearning | VE SDE · MNIST classes 3 and 7 | Unlearning ratio (%): 5.1 -> 1.2 | Improved |
| Use MSGM instead of EraseDiff | VP SDE · CIFAR-10 automobile and dog | Unlearning ratio (%): 17.9 -> 11.9 | Improved |
| Fine-tune a pre-trained DDPM with Obtuse-MSGM instead of EraseDiff | DDPM · CIFAR-10 automobile and dog (fine-tuned) | Unlearning ratio (%): 7.4 -> 5.2 | Improved |
| Use MSGM instead of Unseen in SD v1.4 unlearning | SD v1.4 · Imagenette tench | FID of D_g: 3.09 -> 3.08 | Improved |
| Use pseudo-label cyprinoid with EraseDiff rather than MSGM | EraseDiff · Imagenette tench | Classification rate: 0.00 -> 0.47 | Hurt |
| Use Obtuse-MSGM instead of Standard on inpainting | VP SDE · CelebA inpainting D_f | Classification accuracy (%): 53.0 -> 0.5 | Improved |
| Use Obtuse-MSGM instead of Standard on inpainting | VP SDE · CIFAR-10 inpainting D_f | Classification accuracy (%): 75.0 -> 49.6 | Improved |
Closest papers by meaning
- Algorithm- and Data-Dependent Generalization Bounds for Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Approximation and Generalization Abilities of Score-based Neural Network Generative Models for Sub-Gaussian Distributions (NeurIPS 2025 · similarity 0.90)
- Score Forgetting Distillation: A Swift, Data-Free Method for Machine Unlearning in Diffusion Models (ICLR 2025 · similarity 0.89)
- Beyond Log-Concavity and Score Regularity: Improved Convergence Bounds for Score-Based Generative Models in W2-distance (ICML 2025 · similarity 0.89)
- Score-based generative models are provably robust: an uncertainty quantification perspective (NeurIPS 2024 · similarity 0.88)
- Supervised Score-Based Modeling by Gradient Boosting (AAAI 2025 · similarity 0.88)
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.
