Flash-BoN explores diffusion-model scaling by elapsed time, not evaluation counts

A Flash-BoN team member says the work began with a shift in how diffusion-model inference scaling is measured: elapsed time rather than the number of function evaluations. The aim is to use generation-time compute to explore more candidates. The team member says Flash-BoN works across text-to-image and text-to-video generation and different model scales, complements BFS and ReflectionFlow, and improves Flow-GRPO’s convergence.