Flickering after quantization? A temporally stable 2-bit KV Cache compression framework for video world models emerges

Video world models need to generate frames continuously and maintain scene consistency under the control of user-specified actions. For the autoregressive generation route, the KV Cache stores historical frames, enabling the model to retrieve and reuse previously generated similar scenes during generation. The cost is that memory accumulates continuously over the generation process: when LingBot-World-v2 generates a video of about 5 seconds and 93 frames, the KV Cache alone requires more than 21 GiB of memory. Compressing the KV Cache to 2-bit therefore becomes an important direction for alleviating memory pressure.

The existing method Quant-VideoGen (QVG) has already applied 2-bit KV Cache quantization to autoregressive video generation and achieved performance close to the BF16 model on video benchmarks. But the research team found that when these methods are applied to video world models, even if the benchmark scores are almost unchanged, the generated frames may still show obvious flickering, blurring and detail distortion.

The most intuitive evidence comes from HY-World 1.5: the VBench Temporal flickering scores of BF16 and QVG are 95.62% and 95.85% respectively, with the latter even slightly higher, yet watching the generated videos still reveals severe frame flickering and visual degradation introduced by quantization. This shows that even a benchmark metric directly related to flickering fails to fully capture the quantization performance degradation on video world models.

The KV Cache contains Key and Value. The Key can be understood as the "index" for retrieving historical information, while the Value is the "content" read after retrieval. After quantizing the two separately, the research team found that compressing only the Key to 2-bit brings more obvious visual degradation than quantizing only the Value; more notably, compared with the Value, the Key's own quantization error is often smaller, yet it still causes a larger output error. This means the difference in video quality is not because the Key is harder to quantize; looking only at the magnitude of quantization error cannot explain the temporal flickering and quality degradation.

The research team points the explanation to the attention computation itself. The matching scores between the current Query and historical Keys determine how the model allocates attention to historical information. The error introduced by quantizing the Key may change the relative ranking of these scores, causing the model to attend to another frame or another spatial position instead. For the same Query, the highest-scoring historical spatiotemporal position before and after quantization changes, sometimes even crossing different frames. For continuous generation, this shift may perturb the model's utilization of historical scenes, thereby affecting frame stability.

The judgment drawn from this is: when compressing the KV Cache, one must not only compress the data, but also protect the way the model looks up historical information. This judgment forms the starting point for QuantWM's design. It should be noted that how exactly the shift in the highest-scoring historical spatiotemporal position caused by Key quantization perturbs the model's utilization of historical scenes is still listed in the source as a question requiring further analysis.

The first design is Quantization Sensitivity-Aware Clustering (QSAC). In QVG, each Key is assigned to a cluster center, the center retains higher precision, and the residual between the two is quantized to 2-bit. The research team believes that the nearest center is not necessarily the most suitable result after quantization; the dynamic range of the residual and which channels the error falls on both affect the final attention scores. QSAC first uses a Query-sensitivity-weighted distance to screen candidate centers, then combines the dynamic range of each residual group and its sensitivity to estimate the attention perturbation that INT2 quantization may introduce, and selects the center accordingly.

The second design is Principal Subspace Attention Compensation (PSAC). Even after QSAC, 2-bit quantization inevitably still has errors; the most direct solution is to compensate for the error on the attention logits, but storing the full quantization error is very expensive. The team observed that the energy of historical Queries tends to concentrate in a few directions: for one layer of LingBot-World-v2, just the first 8 principal directions cover about 71% of the Query energy. Based on this, PSAC stores in low-rank form the components of the Key quantization error along these principal directions, and compensates the Key during cache reconstruction, thereby correcting the attention scores computed subsequently. In the experiments, 8 principal directions are used, and the compensation coefficients are quantized to INT8 to control the extra storage overhead.

Both designs use historical Query statistics from already-generated content, require no retraining of the model, and can be directly plugged into the generation pipeline of video world models.

The team validated QuantWM on three video world models: Matrix-Game-2, LingBot-World-v2 and HY-World 1.5, and further validated its generalization ability on two video generation models: LongCat-Video and Causal-Forcing.

In the comparison of 480p, 93-frame generated videos, QuantWM's PSNR, SSIM and LPIPS all outperform QVG and KIVI, and the visualizations show that phenomena such as temporal flickering and blurring are improved. On the five evaluation dimensions of VBench, QuantWM achieves the best average ranking among the compared quantization methods across all models. The paper further reports evaluation results for 720p and 1-minute long videos, validating generation quality at higher resolution and video benchmark performance under long-video generation respectively, but the specific values of these results are not given in the source.

Attention analysis further shows that QuantWM can reduce the shift in historical spatiotemporal position selection caused by quantization: on Matrix-Game-2, the proportion of selection changes in the highest-scoring historical tokens relative to BF16 drops from 53.88% to 10.19%; on HY-World 1.5, it drops from 61.36% to 13.92%.

On the system implementation side, QuantWM designs dedicated Triton kernels; after accounting for extra overhead such as cluster centers, quantization parameters and compensation information, actual KV Cache compression reaches up to 6.20x. It should be pointed out that the above results, including PSNR, SSIM, LPIPS and VBench rankings, are all self-reported by the paper, and the source provides no third-party independent reproduction; the maximum 6.20x compression is a measured value after accounting for extra overhead, and the compression performance under different models and resolutions is not explained one by one in the source.

What QuantWM reveals is an easily overlooked problem in low-bit KV Cache compression for video world models: even if video benchmark scores are close, it does not mean that historical scene information can still be stably utilized after quantization. The comparison of 95.62% and 95.85% on HY-World 1.5 is exactly a typical sample of this "close scores, degraded viewing experience."

The methodological implication extending from this is that compression research for video world models needs multi-faceted evaluation combining benchmark metrics, actual video performance, and the way the model utilizes historical information. This conclusion itself comes from the research team's argument, and the extent to which it can be generalized still depends on subsequent independent verification.