Research map / Image generation / Fast diffusion sampling
DriftLite: Lightweight Drift Control for Inference-Time Scaling of Diffusion Models
Yinuo Ren, Wenhao Gao, Lexing Ying, Grant M. Rotskoff, Jiequn Han
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 VCG drift control with SMC resampling to guidance-SMC | G-SMC · DW-4 annealing | SWD: 1.012 -> 0.035 | Improved |
| add VCG drift control with SMC resampling | G-SMC · LJ-13 annealing | MMD: 0.616 -> 0.102 | Improved |
| add VCG drift control with SMC resampling | G-SMC · LJ-13 reward-tilting | MMD: 0.081 -> 0.015 | Improved |
| add resampling (SMC) to VCG (ablation) | VCG · LJ-13 annealing | energy W1: 22.86 -> 0.286 | Improved |
| add resampling (SMC) to VCG (ablation) | VCG · DW-4 reward-tilting | MMD: 0.614 -> 0.021 | Improved |
| add resampling (SMC) to VCG (ablation) | VCG · DW-4 annealing | MMD: 0.014 -> 0.014 | No clear effect |
| use VCG-SMC to steer Boltz-2 toward physical validity | Boltz-2 base · PoseBusters V2 | Valid Fraction: 0.374 -> 0.856 | Improved |
| use VCG-SMC instead of G-SMC for co-folding steering | G-SMC · PoseBusters V2 | Valid Fraction: 0.838 -> 0.856 | Improved |
| use VCG-SMC instead of FKS for co-folding steering | FKS · PoseBusters V2 | Valid Fraction: 0.379 -> 0.856 | Improved |
| use VCG-SMC instead of G-SMC for internal clashes | G-SMC · PoseBusters V2 | Internal Clash violations: 31.33 -> 32.33 | No clear effect |
| use VCG (no resampling) without SMC on harder tasks (ablation) | VCG · LJ-13 reward-tilting | MMD: 0.719 (PG) -> 0.695 (VCG) | No clear effect |
| increase particle count (ablation) | VCG/ECG · GMM annealing | sample quality | Improved |
Closest papers by meaning
- A Mean-Field Framework for Inference-Time Distributional Control of Diffusion Models (NeurIPS 2026 · similarity 0.89)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.89)
- A General Framework for Inference-time Scaling and Steering of Diffusion Models (ICML 2025 · similarity 0.88)
- Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities (NeurIPS 2025 · similarity 0.88)
- Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · similarity 0.87)
- Transport meets Variational Inference: Controlled Monte Carlo Diffusions (ICLR 2024 · similarity 0.87)
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.
