Coral-KV Compresses Video Model KV-Cache by 37–48% and Lifts PSNR by Up to 4.95 dB

Long video runs into a memory wall, not a compute wall. Autoregressive video diffusion models keep attention over every frame generated so far, so the KV-cache grows with the clip. The Coral-KV preprint on arXiv proposes compressing this cache differently from current methods. On MAGI-1, the scheme delivers +4.95 dB PSNR over QVG-4 with a 37% smaller cache. On HY-WorldPlay the gain is more modest, +0.32 dB, but memory drops by 48%.

Where Cluster-Based Quantization Breaks Down

Until now, clustering-based quantization has been considered the strongest approach to video cache compression. Similar key and value vectors are grouped together, and only the residual errors relative to the cluster centers are quantized. The paper's baselines, QVG-4 and QVG-Pro, belong to this family.

The method has a weak spot, which the authors state plainly: later frames attend to history that has already been compressed. A small reconstruction error in the old cache leaks into new frames, those frames become part of the history, and the error accumulates as the video grows. In short clips this is barely noticeable. In long generation it shows up as quality degradation and temporal discontinuities.

Predicting

Across Layers Instead of Clustering Within a Layer

Clustering exploits similarity within a single layer. The Coral-KV authors found that video models also have strong dependencies between cache layers. That led to the idea: use the cache states in one layer to predict the keys and values of subsequent layers, and quantize only the prediction residual. The more accurate the prediction, the smaller the residual and the fewer bits needed to store it.

The predictors are kept compact, in a block-DPLR format: a block-diagonal matrix plus a low-rank term. The block-diagonal part describes relationships between channels within each attention head. The global low-rank term captures dependencies between heads. This structure is far cheaper than a full transformation matrix between layers.

The name stands for Calibrated Offline, Residual-coded, Adapted Locally. It describes the two operating modes:

• reusable configuration: predictors are calibrated in advance and applied to different videos without further tuning;
• adapted configuration: lightweight block-diagonal predictors are rebuilt for the current video for higher accuracy.

Numbers from the Paper

• HY-WorldPlay, reusable mode vs. QVG-4: +0.32 dB PSNR, cache 48% smaller.
• MAGI-1, reusable mode vs. QVG-4: +4.95 dB PSNR, cache 37% smaller.
• MAGI-1, adapted mode vs. QVG-Pro: +3.77 dB PSNR, cache 33.7% smaller.
• On long generation, the method maintains higher reconstruction accuracy and reduces temporal discontinuities.
• Generation latency stays comparable to the baselines. The abstract gives no specific figures.

The spread between models is notable: on one model the quality gain is almost token, on the other it is huge. Cross-layer redundancy is apparently stronger in MAGI-1, but according to the abstract this is only a guess.

Who Can Use This

For teams running long-context video models, the takeaway is this: KV-cache memory can be cut by a third or more without quality loss if you exploit redundancy between layers. The reusable-predictor mode suits a service with lots of videos, where there is no time to fit predictors for each one. The adapted mode makes sense where quality matters more, such as generating long scenes or interactive worlds, which is what HY-WorldPlay targets.

The caveats are these. This is a first-version preprint, without peer review. Evaluation relies on PSNR, which does not always match perceived video quality. The code, the cost of calibration, and the prediction overhead in real-world serving have yet to be verified. If the results reproduce, cache compression could become one of the most practical ways to extend generated clips on the same hardware.