arXiv：cs.LG· Marcello Bullo, Yanxiao Liu, \"Oyk\"u S{\i}la G\"uner, Arpan Mukherjee, Deniz G\"und\"uz·· 3 小时前AI 评分33
RASS：面向扩散草稿树的排名感知投机采样
Rank-Aware Speculative Sampling for Diffusion Draft Trees
AI 导读
研究者提出 Rank-Aware Speculative Sampling（RASS），一种基于排名感知列表耦合的扩散草稿树验证规则，按提议-目标均值位移对候选排序并采样排名，再与目标最大耦合、以残差校正保证精确采样。在 CIFAR-10、FFHQ、Stable Diffusion 3.5 等评测中，RASS 在同等算力下优于 D-GRS，CIFAR-10 上目标模型评估次数减少约 20%。
正文
Abstract:Speculative sampling accelerates diffusion generation by verifying inexpensive draft states in parallel while preserving the target law. Recent tree-based methods allocate the parallel compute budget more effectively than single-chain drafts, as demonstrated by Diffusion Greedy Rejection Sampling (D-GRS). D-GRS generates $K$ conditionally independent candidates per node, and sequentially tests them in their generation order. Yet the sampled candidates admit an informative ranking without additional target-model evaluations. To exploit this, we introduce Rank-Aware Speculative Sampling (RASS), a verification rule for speculative draft trees based on rank-aware list coupling. RASS orders draft candidates along the proposal-target mean displacement and samples a rank with weights optimized to minimize total variation between the selected-proposal and target laws. Finally, the selected candidate is maximally coupled with the target, with residual correction ensuring exact sampling for any choice of rank weights. We evaluate RASS on a Gaussian-mixture target, unconditional pixel-space generation on FFHQ, conditional generation on CIFAR-10, and latent diffusion with Stable Diffusion 3.5 using COCO2014 prompts. Measured by the ratio of standard to speculative sampling's target-model evaluation counts, RASS improves on D-GRS across the evaluated settings, with gains reaching approximately 20% on CIFAR-10 at matched compute budgets.
| Subjects: | Machine Learning (cs.LG); Computation (stat.CO) |
| Cite as: | arXiv:2610.02251 [cs.LG] |
| (or arXiv:2610.02251v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.02251 arXiv-issued DOI via DataCite |
Submission history
From: Marcello Bullo [view email]
[v1]
Wed, 30 Sep 2026 18:59:32 UTC (14,263 KB)
来源：arXiv：cs.LG · arxiv.org