Research map / Probabilistic and scientific modeling / Variational and generative sampling
Stein Diffusion Guidance: Training-Free Posterior Correction for Sampling Beyond High-Density Regions
Van Khoa Nguyen, Lionel Blondé, Alexandros Kalousis
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 Stein correction to Tweedie-based posterior (ablation) | SDG · Novel hit ratio 5ht1b | novel hit ratio (%): 0.033 -> 22.690 | Improved |
| add Stein correction (ablation) | SDG · Novel hit ratio Jak2 | novel hit ratio (%): 0.000 -> 9.167 | Improved |
| add Stein correction (ablation) | SDG (alpha=0) · Super resolution | FID: 91.99 -> 68.90 | Improved |
| add Stein correction (ablation) | SDG (alpha=0) · Gaussian deblur | FID: 87.23 -> 70.00 | Improved |
| add Stein correction (ablation) | SDG (alpha=0) · Label guidance | accuracy (%): 48.2 -> 54.0 | Improved |
| add Stein correction (ablation) | SDG (alpha=0) · Label guidance | FID: 89.50 -> 105.4 | Hurt |
| add Stein correction (ablation) | SDG (alpha=0) · T2I style transfer | CLIP score: 29.0 -> 28.8 | Hurt |
| add low-density annealing alpha(t)>0 (ablation) | SDG · Novel hit ratio 5ht1b | novel hit ratio (%): 21.278 -> 22.690 | Improved |
| use Stein repulsive correction instead of Langevin (epsilon=0) (ablation) | SDG · Novel hit ratio Parp1 | novel hit ratio (%): 7.933 -> 8.780 | Improved |
| increase particles from 512 to 1024 to 3000 (ablation) | SDG · Novel hit ratio 5ht1b | novel hit ratio (%): 22.690 -> 21.922 -> 22.389 | No clear effect |
| add Stein correction on molecule guidance (compute cost) (ablation) | SDG · Molecule guidance runtime | runtime (s): 1360 -> 2521 | Hurt |
| SDG versus MOOD classifier guidance | GDSS · Novel hit ratio Jak2 | novel hit ratio (%): 9.200 -> 9.167 | No clear effect |
Closest papers by meaning
- Multilevel and Sequential Monte Carlo for Training-Free Diffusion Guidance (NeurIPS 2026 · similarity 0.89)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.89)
- Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · similarity 0.89)
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.88)
- Simple Guidance Mechanisms for Discrete Diffusion Models (ICLR 2025 · similarity 0.88)
- Feedback Guidance of Diffusion Models (NeurIPS 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.
