Flash-BoN: cheap drafts plus staged verification for diffusion scaling
RisingSayak · x · 2026-09-11
The author will present three image-generation works at ECCV 2026 posters, including dynamic eval systems and inference-time scaling in diffusion/flow-matching, with Flash-BoN addressing BoN sampling costs.
Flash-BoN highlights:
- NFE-based comparisons ignore verifier overhead; under wall-clock budgets, simple BoN already matches or beats several guided search methods.
- It generates a large pool of cheap drafts via timestep truncation, layer skipping, and activation proxies (optimized once per model), then multi-stage verification picks the best draft for full-quality refinement.
- It consistently outperforms baselines across three benchmarks and three model scales, with gains growing at scale (+8% AUC), and combines well with orthogonal techniques.
More from Multimodal
- Creator builds Snorricam video series with Seedance 2.5, sharing prompts — LudovicCreator · 2026-10-27
- Gradium launches free webpage TTS Chrome extension powered by its 40M-parameter Phonon model — mattturck · 2026-10-27
- Vortex of Infinite Recursion prompt goes viral: images that spiral endlessly into themselves — azed_ai · 2026-10-27
- 170+ AI Creators Made One Collective Film, Premiering November 6 — Loo_Atreides · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Midjourney creator showcases --sref style reference outputs: gothic character art — OVolosin82152 · 2026-10-27