Research map / Image generation / Fast diffusion sampling
Inference-time scaling of diffusion models through classical search
XiangCheng Zhang, Haowei Lin, Haotian Ye, James Zou, Jianzhu Ma, Yitao Liang, Yilun Du
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 SSP resampling instead of multinomial (ablation) | BFS tree search · SD v1.5 ImageReward, N=4 | ImageReward score: 0.743 -> 0.834 | Improved |
| Use Increase tempering instead of constant (ablation) | BFS tree search · SD v1.5 ImageReward, N=4 | ImageReward score: 0.834 -> 0.882 | Improved |
| Use infinite tempering (top particle only) instead of constant (ablation) | BFS tree search · SD v1.5 ImageReward, N=4 | ImageReward score: 0.834 -> 0.667 | Hurt |
| Use Max scoring instead of Current (ablation) | BFS tree search · SD v1.5 ImageReward, N=4 | ImageReward score: 0.812 -> 0.834 | Improved |
| Use improved BFS instead of prior particle-based methods | BFS · SDXL ImageReward, N=8 | ImageReward score: DAS 1.265 -> 1.291 | Improved |
| Replace BFS with adaptive DFS with backtracking | DFS · CompBench T2I | accuracy vs compute | Improved |
| Scale local Langevin search steps together with global search | Best-of-N / BFS / DFS · PointMaze planning | success vs compute | Improved |
| Use local search only (TFG) without global search | TFG · PointMaze planning | success vs compute | Hurt |
| Use test-time search with local and global search instead of TFG and DAS | TTS · D4RL Medium Hopper | normalized score: DAS 71.3 -> 99.5 | Improved |
| Compare training-free TTS with training-based QGPO | TTS · D4RL locomotion average | normalized score: QGPO 86.6 vs TTS 86.1 | No clear effect |
| Distill TTS actions into diffusion policy | TTS-distill · D4RL Medium Hopper | normalized score: DPPO 92.8 -> 98.8 | Improved |
| Use separate verifiers for local and global search (double verifier) (ablation) | BFS · ImageNet conditional, N=8 | FID: 133.3 -> 118.2 | Improved |
Closest papers by meaning
- Adaptive Inference-Time Scaling via Cyclic Diffusion Search (NeurIPS 2025 · similarity 0.91)
- Diffusion Tree Sampling: Scalable inference‑time alignment of diffusion models (NeurIPS 2025 · similarity 0.89)
- Test-time Alignment of Diffusion Models without Reward Over-optimization (ICLR 2025 · similarity 0.89)
- Hierarchical Variational Policies for Reward-Guided Diffusion (NeurIPS 2026 · similarity 0.89)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.89)
- Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing (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.
