Research map / Probabilistic and scientific modeling / Variational and generative sampling
Simple Approximation and Derivative Free Inference-Time Scaling for Diffusion Models via Sequential Monte Carlo on Path Measures
Chenyang Wang, Weizhong Wang, Yinuo Ren, Jose Blanchet, Yiping Lu
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 |
|---|---|---|---|
| use pathwise Girsanov weights (URGE) instead of FK-Steering reward-difference weights | SMC guidance on GMM · 30-d GMM | Cov Frobenius error: 198.20 -> 181.31 | Improved |
| use derivative-free URGE instead of AFDPS+VCG | GMM guidance · 30-d GMM | MMD: 0.08 -> 0.06 | Improved |
| add resampling on top of pure guidance (AFDPS) | Pure guidance · 30-d GMM | MMD: 0.17 -> 0.10 | Improved |
| add variance-controlling guidance (VCG) to AFDPS | AFDPS · 30-d GMM | MMD: 0.10 -> 0.08 | Improved |
| increase number of particles for URGE (ablation) | URGE · 30-d GMM | accuracy and runtime | Improved |
| use URGE instead of AFDPS variants on inverse problems | ImageNet-256 inverse problems · ImageNet-256 inverse problems | PSNR/LPIPS: 22.38 vs 22.43 (Gaussian deblurring PSNR) | No clear effect |
| use URGE instead of AFDPS-ODE on motion deblurring | ImageNet-256 motion deblurring · ImageNet-256 motion deblurring | PSNR: 21.46 -> 18.37 | Hurt |
| use URGE instead of FK-Corrector on inverse problems | ImageNet-256 Gaussian deblurring · ImageNet-256 Gaussian deblurring | PSNR: 18.36 -> 22.38 | Improved |
| use URGE (N=4) instead of FK-Steering (N=4) | SD v1.5 text-to-image · 50 prompts | ImageReward: 0.8397 -> 0.9955 | Improved |
| use URGE (N=4) instead of single sample N=1 | SD v1.5 text-to-image · 50 prompts | GenEval: 0.6400 -> 0.7800 | Improved |
| add gradient guidance to FK-Steering (ablation) | FK-Steering N=4 · 50 prompts | ImageReward: 0.8397 -> 0.7906 | Hurt |
| add gradient guidance to the N=1 sampler (ablation) | SD v1.5 N=1 · 50 prompts | ImageReward: 0.2144 -> 0.2074 | No clear effect |
Closest papers by meaning
- Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing (NeurIPS 2025 · similarity 0.91)
- A General Framework for Inference-time Scaling and Steering of Diffusion Models (ICML 2025 · similarity 0.91)
- Diffusion Tree Sampling: Scalable inference‑time alignment of diffusion models (NeurIPS 2025 · similarity 0.91)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.91)
- Accelerating Convergence of Score-Based Diffusion Models, Provably (ICML 2024 · similarity 0.90)
- A Mean-Field Framework for Inference-Time Distributional Control of Diffusion Models (NeurIPS 2026 · similarity 0.90)
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.
