[论文] Speculative Forcing for Efficient Autoregressive Video Generation

论文概要

研究领域: CV
作者: Hyunwook Choi, Seungryong Kim, Dongyoon Han, Sangdoo Yun
发布时间: 2026-10-15
arXiv: 2610.22347

中文摘要

少步自回归视频生成器实现了流式生成，但它们让每个 token 都经过最终去噪步骤，无论需要多少细化。我们发现这种统一计算通常是不必要的：许多 token 在初始步骤中已与其最终输出非常接近，而其他需要直到最后都在细化。我们引入 Speculative Forcing，将中间干净估计视为草稿并预测何时可以安全提前接受单个 token。一旦被接受，token 就被冻结为最终输出，不再进行去噪，其缓存的键和值仍可作为剩余 token 的上下文。在 LongLive 上，它以每个 token 约 1.7 个去噪步骤匹配四步生成器，运行速度超过 1.5 倍，达到 41.7 和 37.8 fps。在 Lingbot-World 2.0 上，它以约 2.5 个去噪步骤保持质量，运行速度快 1.35 倍。

原文摘要

Few-step autoregressive video generators enable streaming generation, yet they run every token through the final denoising step, regardless of how much refinement it needs. We find this uniform computation is often unnecessary: many tokens already closely match their eventual full-step outputs in the initial steps, while others require refinement until the end. We introduce Speculative Forcing, a framework that treats intermediate clean estimates as drafts and predicts when individual tokens can be safely accepted early. Once accepted, a token is frozen as its final output and undergoes no further denoising, while its cached keys and values remain available as context for the remaining tokens. We explore two alternative verifier designs: a training-free one based on similarity to clean con...

自动采集于 2026-10-17

#论文 #arXiv #CV #小凯