Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models

This paper reexamines inference-time scaling for diffusion models under wall-clock budgets. It reports that simple Best-of-N sampling can match or outperform several guided-search methods once verifier overhead is counted, because those methods spend substantial compute on intermediate checks. Flash-BoN creates many inexpensive draft candidates by combining timestep truncation, layer skipping, and activation proxies in one model-specific configuration, then verifies and fully refines the best candidate. Across three benchmarks and three model scales, it reportedly outperforms all baselines under fixed wall-clock budgets, with gains reaching +8% AUC at larger scales. It also improves reflection-based prompt optimization by +16% AUC and supports faster RL post-training convergence through greater candidate diversity.