Flash-BoN Says Wall-Clock Time Flips Diffusion Rankings

A small methodology paper on arXiv this week quietly argues that most of the diffusion inference-scaling leaderboards have been ranking the wrong thing. The claim from Rawal and coauthors is that when you compare inference-time scaling methods by the number of function evaluations you count only denoising forward passes and ignore the verifier calls that guided search methods lean on. Rerun the same comparison on actual elapsed time and, as they put it, simple BoN already matches or outperforms several guided search techniques. Several of the winners on the old chart go away.

Their alternative, Flash-BoN, spends the compute on more shots on goal rather than more inspection between shots. It bundles three tricks (timestep truncation, layer skipping, and activation proxies) into a single configuration that generates inexpensive draft candidates, cheaply picks the promising one, and only spends full-quality compute on the finalist. The authors report gains across GenAI-Bench, GenEval, and UniGenBench, tested on Wan2.1 at 1.3B and 14B and on FLUX.1-dev, with about +8% AUC at the larger scales and about +16% AUC when they layer in reflection-based prompt optimization. Draft-pool diversity correlates with final quality at Pearson r=0.75, which is their argument that broader exploration delivers different solutions rather than just more lottery tickets.

Why this matters if you actually run an image-gen stack: if your throughput and cost model was built around guided search or repeated intermediate verification, the reported wall-clock inversion says you may have been optimizing for a metric that flatters that architecture. The other line worth flagging is that they fold the same draft pattern into RL post-training as Flash-Flow-GRPO and report baseline convergence in roughly 10x fewer gradient steps, which changes the cost of running the fine-tune loop, not just serving.

The honest caveats. It is one paper, on three benchmarks, on open-weight models, and the +8% and +16% figures are AUC gains rather than human-preference win rates on production prompts. What the reporting does not give you is whether the ranking inversion still holds against the newest closed-source image models, or whether truncation and layer skipping cost quality on the long, structured prompts that break diffusion models today. Take the specifics as reported, not settled.

Still, the direction is the interesting part. If the community shifts to wall-clock reporting instead of NFEs, the incentive to build ever more elaborate intermediate verifiers softens and the incentive to make the draft path faster gets stronger. That is a better fit for anyone who actually pays the latency bill.