CORAL-KV: Compact Cross-Layer Prediction and Residual Coding for Autoregressive Video KV Caches

Computer Science > Computer Vision and Pattern Recognition arXiv:2610.12514 (cs) [Submitted on 7 Oct 2026] Title: CORAL-KV: Compact Cross-Layer Prediction and Residual Coding for Autoregressive Video KV Caches Authors: Albert Ge, Akash Haridas, Utkarsh Saxena, Ruijing Yang, Ian Colbert, Frederic Sala, Vikram Appia, Emad Barsoum.

计算机科学 > 计算机视觉与模式识别 arXiv:2610.12514 (cs) [提交于 2026 年 10 月 7 日] 标题：CORAL-KV：用于自回归视频 KV 缓存的紧凑跨层预测与残差编码。作者：Albert Ge, Akash Haridas, Utkarsh Saxena, Ruijing Yang, Ian Colbert, Frederic Sala, Vikram Appia, Emad Barsoum。

Abstract: Autoregressive video diffusion models maintain a growing attention key-value (KV) cache, making memory a key bottleneck for long video generation. Recently, clustering-based quantization has emerged as an effective method to compress video KV caches by grouping similar key and value vectors then quantizing residual errors. However, when later frames attend to the compressed history, small reconstruction errors can compound as the video grows.

摘要：自回归视频扩散模型维护着一个不断增长的注意力键值（KV）缓存，这使得内存成为长视频生成的关键瓶颈。最近，基于聚类的量化已成为压缩视频 KV 缓存的有效方法，通过对相似的键和值向量进行分组，然后量化残差误差。然而，当后续帧关注压缩后的历史记录时，随着视频的增长，微小的重建误差可能会累积。

Beyond the within-layer similarity exploited by clustering, we find strong dependencies across cache layers in video models. This motivates Coral-KV - Calibrated Offline, Residual-coded, and Adapted Locally - a KV-cache compression framework that uses cache states in each layer to predict keys and values in subsequent layers, then quantizes the remaining prediction residuals.

除了聚类所利用的层内相似性之外，我们发现视频模型中各缓存层之间存在很强的依赖关系。这促成了 Coral-KV（离线校准、残差编码和局部自适应）的诞生，这是一个 KV 缓存压缩框架，它利用每一层的缓存状态来预测后续层的键和值，然后对剩余的预测残差进行量化。

We construct compact block-diagonal-plus-low-rank (block-DPLR) predictors: block-diagonal components model channel relationships within each attention head, while a global low-rank term models dependencies between heads. Coral-KV exposes a flexible fidelity-memory trade-off: it can reuse its predictors across videos or refit lightweight block-diagonal predictors to the current video for higher fidelity.

我们构建了紧凑的块对角加低秩（block-DPLR）预测器：块对角分量用于建模每个注意力头内的通道关系，而全局低秩项则用于建模各头之间的依赖关系。Coral-KV 提供了一种灵活的保真度与内存权衡方案：它可以在不同视频间重用预测器，或者为当前视频重新拟合轻量级的块对角预测器以获得更高的保真度。

Across extensive evaluations on both HY-WorldPlay and MAGI-1, our reusable configuration improves PSNR over QVG-4 by 0.32 dB and 4.95 dB, respectively, while using 48% and 37% less cache memory. Against QVG-Pro, our video-adapted configuration improves MAGI-1 PSNR by 3.77 dB while using 33.7% less cache memory. Coral-KV also maintains higher reconstruction fidelity over extended generation and reduces temporal discontinuities while retaining comparable generation latency.

在对 HY-WorldPlay 和 MAGI-1 进行的广泛评估中，我们的可重用配置使 PSNR 分别比 QVG-4 提高了 0.32 dB 和 4.95 dB，同时缓存内存使用量分别减少了 48% 和 37%。与 QVG-Pro 相比，我们的视频自适应配置在保持 33.7% 更少缓存内存的同时，将 MAGI-1 的 PSNR 提高了 3.77 dB。Coral-KV 在长时间生成过程中也保持了更高的重建保真度，并减少了时间上的不连续性，同时保留了相当的生成延迟。