Flash-BoN presented at ECCV 2026, focusing on elapsed time for diffusion scaling

The team presented Flash-BoN at ECCV 2026 on September 11, according to a researcher on the project. The researcher describes using wall-clock time—actual elapsed time—as the metric for scaling diffusion inference, rather than counting function evaluations. They say Flash-BoN works for text-to-image and text-to-video generation across model scales, complements BFS and ReflectionFlow, and improves Flow-GRPO’s convergence.