Flash-BoN: Rethinking Inference-Time Scaling for Diffusion Models

Key Info

Hugging Face researcher Sayak Paul presented Flash-BoN at ECCV, a project that explores inference-time scaling for diffusion models by shifting the optimization metric from number of function evaluations (NFE) to wall-clock time.

Highlights

• Flash-BoN was presented on September 11 at ECCV 2026, with origins tracing back to a conversation at ICCV 2025 in Hawai'i.
• The work proposes using more bounded metrics like wall-clock time to allocate inference-time compute more effectively.
• It highlights a growing push to make inference-time scaling for diffusion models practical and efficient.