Research map / Image generation / Fast diffusion sampling
Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models
Ruchit Rawal, Reza Shirkavand, Sayak Paul, Yuxin Wen, Heng Huang, Yizheng Chen, Tom Goldstein, Gowthami Somepalli
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 optimized draft pool (Flash-BoN) instead of plain BoN under wall-clock budget | Best-of-N · GenAI/GenEval/UniGenBench, 3 models | AUC/Time: e.g. 0.60 -> 0.68 (Wan 14B GenAI-Bench); 0.65 -> 0.66 (Wan 1.3B) | Improved |
| use guided search (BFS/DFS/ZOS) instead of BoN under wall-clock budget | BoN · GenAI-Bench, Wan2.1 14B | AUC/Time: BoN 0.60 vs BFS 0.43, DFS 0.50, ZOS 0.57 | Hurt |
| add layer skipping to timestep truncation drafts (ablation) | Flash-BoN drafts · GenAI-Bench, Wan2.1 1.3B | AUC: 0.623 -> 0.646 | Improved |
| add activation proxies to TS+LS (ablation) | Flash-BoN drafts · GenAI-Bench, Wan2.1 1.3B | AUC: 0.646 -> 0.652 | Improved |
| add multi-stage pairwise verification to three knobs (ablation) | Flash-BoN · GenAI-Bench, Wan2.1 1.3B | AUC: 0.652 -> 0.666 | Improved |
| use random draft configuration instead of optimized configuration (ablation) | Flash-BoN · GenAI-Bench, Wan2.1 1.3B | AUC: 0.666 -> 0.551 | Hurt |
| use FLUX.1-dev draft config on Wan2.1 14B instead of native calibration (ablation) | Flash-BoN · GenAI-Bench, Wan2.1 14B | AUC: 0.68 -> 0.61 | Hurt |
| layer Flash drafts onto Reflection-Tuning | Reflection-Tuning · GenAI-Bench | AUC: 0.46 -> 0.62 | Improved |
| layer Flash drafts onto BFS | BFS · GenAI-Bench | AUC: 0.49 -> 0.55 | Improved |
| replace random rollouts with draft-guided selection (16 drafts, top 6 + bottom 2) | Flow-GRPO · Wan2.1 1.3B RL post-training | reward: 0.692 -> 0.699; matches baseline final at step 60 | Improved |
| use larger 72B verifier instead of 7B (ablation) | Flash-BoN · GenAI-Bench, Wan2.1 1.3B | score under wall-clock | Hurt |
| use 3B verifier instead of 7B (ablation) | Flash-BoN · GenAI-Bench, Wan2.1 1.3B | score under wall-clock | Hurt |
Closest papers by meaning
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.91)
- Scale Where It Matters: Training-Free Localized Scaling for Diffusion Models (NeurIPS 2026 · similarity 0.90)
- Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing (NeurIPS 2025 · similarity 0.90)
- Test-Time Scaling of Diffusion Models via Noise Trajectory Search (NeurIPS 2025 · similarity 0.90)
- SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer (ICML 2025 · similarity 0.90)
- Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models (NeurIPS 2025 · similarity 0.89)
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.
