Flash-BoN at ECCV 2026: rethinking diffusion inference-time scaling with wall-clock time as the budget
RisingSayak · x · 2026-09-13
The Flash-BoN paper was presented at ECCV 2026. The core idea: replace number of function evaluations (NFE) with wall-clock time as the metric for diffusion inference-time scaling — spending bounded time exploring more candidates. It works broadly: across T2I and T2V, different model scales, complements techniques like BFS and ReflectionFlow, and even improves Flow-GRPO convergence. The idea originated from a conversation with Ruchit Rawal at ICCV'25 in Hawai'i.
More from Multimodal
- CTC-TTS paper replaces MFA with CTC alignment for low-latency dual-streaming TTS — kastnerkyle · 2026-09-13
- I Set Out to Test If AI Video Could Sustain a Feature Film and Made Seven — bdylsing · 2026-09-13
- Seedance 2.5 demo: one long prompt generates a cinematic school stealth short — SimplyAnnisa · 2026-09-13
- Creator shares full Seedance prompt for large-scale anime battle scenes — LudovicCreator · 2026-09-13
- Single-word prompting: 'Snowbroth' shows how obscure literary words steer Midjourney — tisch_eins · 2026-09-13
- Untrained Qwen3.8 Flash Next Generates Impressive Frog-on-Whale SVG Locally — ludos1978 · 2026-09-13