Research map / Image generation / Fast diffusion sampling
General and Efficient Steering of Unconditional Diffusion Models
Qingsong Wang, Mikhail Belkin, Yusu Wang
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 high-noise noise alignment to RFM-only steering (ablation) | NA-RFM · CIFAR-10 label guidance | accuracy: 94.8% -> 96.6% | Improved |
| add high-noise noise alignment to RFM-only steering (ablation) | NA-RFM · CIFAR-10 label guidance | FID: 40.3 -> 41.4 | Hurt |
| add RFM activation steering to noise alignment (ablation) | NA-RFM on SiT-XL/2 · ImageNet (SiT-XL/2) | accuracy: 12.9% -> 61.3% | Improved |
| add RFM activation steering to noise alignment (ablation) | NA-RFM on SiT-XL/2 · ImageNet (SiT-XL/2) | FID: 220.9 -> 151.8 | Improved |
| increase noise alignment strength lambda from 3 to 8 (ablation) | noise alignment only · CIFAR-10 label guidance | accuracy: 62% -> 80% | Improved |
| increase noise alignment strength lambda from 3 to 8 (ablation) | noise alignment only · CIFAR-10 label guidance | FID: 99 -> 120 | Hurt |
| train RFM direction at high noise sigma=5.0 instead of sigma in [0.6, 2.0] (ablation) | NA-RFM · CIFAR-10 | generation accuracy | Hurt |
| replace TFG-4 with NA-RFM | CIFAR-10 label guidance · CIFAR-10 label guidance | accuracy: 77.1% -> 96.6% | Improved |
| replace classifier guidance with NA-RFM | CIFAR-10 label guidance · CIFAR-10 label guidance | accuracy: 86.0% -> 96.6% | Improved |
| replace TFG-4 with NA-RFM | ImageNet 256 · ImageNet 256 | accuracy: 59.8% -> 75.8% | Improved |
| replace TFG-4 with NA-RFM | Birds-525 · Birds-525 fine-grained | accuracy: 2.2% -> 14.1% | Improved |
| replace TFG with NA-RFM | CelebA-HQ Gender+Hair · CelebA-HQ Gender+Hair | accuracy: 75.4% -> 83.3% | Improved |
Closest papers by meaning
- Specificity-Aware Diffusion Steering via Variance-Reduced Sequential Monte Carlo (NeurIPS 2026 · similarity 0.88)
- Variational Control for Guidance in Diffusion Models (ICML 2025 · similarity 0.87)
- No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.87)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.86)
- Token Perturbation Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.86)
- Multilevel and Sequential Monte Carlo for Training-Free Diffusion Guidance (NeurIPS 2026 · 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.
