DRiffusion：草稿-精炼过程轻松并行化扩散模型
DRiffusion: Draft-and-Refine Process Parallelizes Diffusion Models with Ease
通过 skip transition 并行生成未来时间步的草稿状态再统一去噪精炼，是一种与具体扩散骨干解耦的并行采样加速手段，可直接套用到图像/视频扩散模型的推理管线中以降低延迟。
原文依据 DRiffusion employs skip transitions to generate multiple draft states for future timesteps and computes their corresponding noises in parallel
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论扩散模型在生成高保真内容方面取得了显著成功，但其迭代式采样速度缓慢，导致高延迟，限制了其在交互式应用中的使用。我们提出 DRiffusion，一个通过草稿-精炼过程并行化扩散推理的并行采样框架。DRiffusion 采用跳跃转移为未来时间步生成多个草稿状态，并并行计算其对应的噪声，随后将这些噪声用于标准去噪过程以产生精炼结果。在理论上，我们的方法可实现 1/n 或 2/(n+1) 的加速比，具体取决于使用保守模式还是激进模式，其中 n 表示设备数量。在实验上，DRiffusion 在多个扩散模型上取得 1.4 倍至 3.7 倍的加速，同时生成质量的下降极小：在 MS-COCO 数据集上，弗雷歇初始距离与 CLIP 指标基本与原模型持平，而 PickScore 与 HPSv2.1 仅分别出现 0.17 和 0.43 的平均小幅下降。这些结果验证了 DRiffusion 在提供显著加速的同时保持了感知质量。
英文原文摘要 1179 字符
Diffusion models have achieved remarkable success in generating high-fidelity content but suffer from slow, iterative sampling, resulting in high latency that limits their use in interactive applications. We introduce DRiffusion, a parallel sampling framework that parallelizes diffusion inference through a draft-and-refine process. DRiffusion employs skip transitions to generate multiple draft states for future timesteps and computes their corresponding noises in parallel, which are then used in the standard denoising process to produce refined results. Theoretically, our method achieves an acceleration rate of \tfrac 1 n or \tfrac 2 n+1 , depending on whether the conservative or aggressive mode is used, where n denotes the number of devices. Empirically, DRiffusion attains 1.4x-3.7x speedup across multiple diffusion models while incur minimal degradation in generation quality: on MS-COCO dataset, both FID and CLIP remain largely on par with those of the original model, while PickScore and HPSv2.1 show only minor average drops of 0.17 and 0.43, respectively. These results verify that DRiffusion delivers substantial acceleration and preserves perceptual quality.