High computational costs and slow inference hinder the practical application of video generation models. While prior works accelerate the generation process through feature caching, they often suffer from notable quality degradation. In this work, we reveal that this issue arises from their inability to distinguish truly redundant features, which leads to the unintended skipping of computations on important features. To address this, we propose PreciseCache, a plug-and-play framework that precisely detects and skips truly redundant computations, thereby accelerating inference without sacrificing quality. Specifically, PreciseCache contains two components: LFCache for step-wise caching and BlockCache for block-wise caching. For LFCache, we compute the Low-Frequency Difference (LFD) between the prediction features of the current step and those from the previous cached step. Empirically, we observe that LFD serves as an effective measure of step-wise redundancy, accurately detecting highly redundant steps whose computation can be skipped through reusing cached features. To further accelerate generation within each non-skipped step, we propose BlockCache, which precisely detects and skips redundant computations at the block level within the network. Extensive experiments on various backbones demonstrate the effectiveness of our PreciseCache, such as achieving an average of 2.6× speedup on Wan2.1-14B without noticeable quality loss.
Figures & tables
Figure 1: Qualitative Results of PreciseCache on Wan2.1-14B ( Wang et al., 2025 ) . Compared with previous methods, our PreciseCache achieves higher acceleration (about 2.6 × speedup) of the base model without sacrificing the quality of generated videos.
Figure 2: The Illustration of the Denoising Process for Video Generation . At high-noise timesteps, the prediction of the model varies significantly. Reusing the cached features in this stage (the red line ) can significantly affect both the content and the quality of generated videos compared to the videos generated without caching (the orange line ). In contrast, the feature caching during low-noise timesteps only introduces negligible impacts (the green line ).
Figure 3: The Impact of Reusing the Cached Model Prediction at Each Timestep. Considering a 50-step denoising process from the Gaussian Noise Z50 to a clean latent Z0 , we respectively reuse the model output at timestep ti+1 for each i∈{49,48,⋯,0} , and perform the subsequent denoising steps to generate the final video. We then compare each resulting video with the baseline (i.e., generated without caching and reusing) to evaluate the impact of reusing cached predictions. Different colors indicate different prompts.
Figure 4: The Difference between Model Predictions at Adjacent Timesteps. Although the difference is relatively large in the low-noise stage, it primarily arises from high-frequency components, which have limited influence on the perceptual quality of the generated results. Different colors indicate different prompts.
Figure 5: Pipeline of PreciseCache.
Figure 6: The Relationship between Latent Resolution and LFD. We observe that downsampling has little effect on the computation of LFD. The experiment is conducted on Wan2.1-14B ( Wang et al., 2025 ) .
Figure 7: The Importance of Each Transformer Block within the Video Diffusion Transformer. Different colors represent different denoising steps. The experiment is conducted on Wan2.1-14B ( Wang et al., 2025 ) .
Method
Efficiency
Visual Quality
MACs (P) ↓
Speedup ↑
Latency (s) ↓
VBench ↑
LPIPS ↓
SSIM ↑
PSNR ↑
Open-Sora 1.2 (480P, 192 frames)
Open-Sora 1.2 ( T=30 )
6.30
1×
47.23
78.79%
-
-
-
PAB
5.33
1.26×
38.40
78.15%
0.1041
0.8821
26.43
TeaCache
3.29
1.95×
24.73
78.23%
0.0974
0.8897
26.84
FasterCache
4.13
1.67×
29.15
78.46%
0.0835
0.8932
27.03
Table 1: Quantitative Comparison of efficiency and visual quality on 4 A800 GPUs.
Figure 8: Qualitative Comparison . Zoom in for the best views.
Table 10
Strategy
VBench ↑
LPIPS ↓
Reuse prediction ( F )
83.52%
0.0793
Reuse residuals ( R )
83.50%
0.0791
TaylorSeer
83.54%
0.0801
Table 4: Feature Reusing Strategy for Step-wise Caching. Without loss of generality, we conduct experiments on Wan2.1-14B ( Wang et al., 2025 ) , generating videos with 1080P resolution.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 9: More Qualitative Results of PreciseCache on Wan2.1-14B . Zoom in for the best views.
Autoregressive video generation paradigms offer theoretical promise for long video synthesis, yet their practical deployment is hindered by the computational burden of sequential iterative denoising. While cache reuse strategies can accelerate generation by skipping redundant denoising steps, existing methods rely on coarse-grained chunk-level skipping that fails to capture fine-grained pixel dynamics. This oversight is critical: pixels with high motion require more denoising steps to prevent error accumulation, while static pixels tolerate aggressive skipping. We formalize this insight theoretically by linking cache errors to residual instability, and propose MotionCache, a motion-aware cache framework that exploits inter-frame differences as a lightweight proxy for pixel-level motion characteristics. MotionCache employs a coarse-to-fine strategy: an initial warm-up phase establishes semantic coherence, followed by motion-weighted cache reuse that dynamically adjusts update frequencies per token. Extensive experiments on state-of-the-art models like SkyReels-V2 and MAGI-1 demonstrate that MotionCache achieves significant speedups of 6.28× and 1.64× respectively, while effectively preserving generation quality (VBench: 1%↓ and 0.01%↓ respectively). The code is available at https://github.com/ywlq/MotionCache.
Jing Xu, Yuexiao Ma, Xuzhe Zheng +7
1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China · 2ByteDance · 3Max Planck Institute for Intelligent Systems, 4ELLIS Institute Tbingen, 5Tbingen AI Center
Feature caching has become a mainstream training-free strategy for accelerating video diffusion models by reusing intermediate outputs across denoising steps. However, existing adaptive methods rely on heuristic difference measures misaligned with actual generation quality, employ a single global threshold that ignores heterogeneous phase dynamics, and never validate their budget estimates against true approximation error, allowing systematic bias to propagate uncorrected. We propose AlignCache, a closed-loop caching framework that directly aligns scheduling decisions with their true influence on the final generated video through three tightly coupled components: (1) phase-adaptive budget routing that partitions the denoising trajectory into structurally distinct phases, each governed by an independent threshold and a staleness-aware guard; (2) closed-loop threshold calibration that exploits counterfactual error measurement at the classifier-free guided output level to adjust decision boundaries online via an EMA feedback loop; and (3) momentum-augmented draft synthesis combining direction-magnitude gated residual extrapolation with anchor-guided unconditional correction. By transforming scheduling from a static risk accumulator into a feedback-driven, self-correcting controller, AlignCache consistently achieves superior acceleration-quality trade-offs across state-of-the-art video generation models. Our code is available at https://github.com/jiangjyjy/AlignCache.
Jiyue Jiang, Zhiyu Huo, Yuhan Chen +3
Department of Computer Science and Engineering, The Chinese University of Hong Kong · Theory Lab, 2012 Labs, Huawei Technology Co. Ltd
Current video generation models suffer from high computational latency, making real-time applications prohibitively costly. In this paper, we address this limitation by exploiting the temporal redundancy inherent in video latent patches. To this end, we propose the Latent Inter-frame Pruning with Attention Recovery (LIPAR) framework, which detects and skips recomputing duplicated latent patches. Additionally, we introduce a novel Attention Recovery mechanism that approximates the attention values of pruned tokens, thereby removing visual artifacts arising from naively applying the pruning method. Empirically, our method increases video editing throughput by 1.53×, achieving an average of 19.3 FPS on an NVIDIA RTX 4090 with the 1.3B Self-Forcing model (4-step denoising, FP16). The proposed method does not compromise generation quality and can be seamlessly integrated with the model without additional training. Our approach effectively bridges the gap between traditional compression algorithms and modern generative pipelines.