WorldCache: Content-Aware Caching for Accelerated Video World Models

Umair Nawaz; Ahmed Heakl; Ufaq Khan; Abdelrahman Shaker; Salman Khan; Fahad Shahbaz Khan

cs.CVMar 23, 2026

WorldCache: Content-Aware Caching for Accelerated Video World Models

Authors: Umair Nawaz, Ahmed Heakl, Ufaq Khan, Abdelrahman Shaker, Salman Khan, Fahad Shahbaz Khan

Organizations: Mohamed bin Zayed University of Artificial Intelligence, UAE · Linköping University, Sweden

Abstract

Diffusion Transformers (DiTs) power high-fidelity video world models but remain computationally expensive due to sequential denoising and costly spatio-temporal attention. Training-free feature caching accelerates inference by reusing intermediate activations across denoising steps; however, existing methods largely rely on a Zero-Order Hold assumption i.e., reusing cached features as static snapshots when global drift is small. This often leads to ghosting artifacts, blur, and motion inconsistencies in dynamic scenes. We propose WorldCache, a Perception-Constrained Dynamical Caching framework that improves both when and how to reuse features. WorldCache introduces motion-adaptive thresholds, saliency-weighted drift estimation, optimal approximation via blending and warping, and phase-aware threshold scheduling across diffusion steps. Our cohesive approach enables adaptive, motion-consistent feature reuse without retraining. On Cosmos-Predict2.5-2B evaluated on PAI-Bench, WorldCache achieves \textbf{2.3 $\times$ } inference speedup while preserving 99.4% of baseline quality, substantially outperforming prior training-free caching approaches. Our code can be accessed on World-Cache.

Figures & tables

Figure 1: Qualitative and quantitative comparison of acceleration methods on video world model generation using Cosmos-Predict2.5-2B. Left: Visual comparison of Baseline (no acceleration), FasterCache, DiCache, and WorldCache (ours) across three representative timesteps ( $T_{1},\ldots,T_{N}$ ) of a driving scene from the City Street domain. FasterCache achieves $1.6\times$ speedup but introduces severe visual artifacts and scene hallucinations. DiCache ( $1.3\times$ ) better preserves scene fidelity but exhibits noticeable spatial artifacts at later timesteps (red dashed boxes). WorldCache (ours) achieves $2.3\times$ speedup while faithfully reproducing scene content, motion, and background structure across all timesteps (green dashed boxes). Right: Relative latency reduction (%) over Baseline for Text2World (T2W) and Image2World (I2W) tasks, showing WorldCache achieves up to $55.5\%$ reduction compared to $40.5\%$ for FasterCache and $27.9\%$ for DiCache. All results are reported on a single H200 GPU.

Figure 2: A DiT world model denoises latent video states ( $z_{t}$ ) by running probe blocks (1 $\ldots$ k) followed by deep blocks ( $k{+}1\ldots(N)$ ). WorldCache inserts a decision gate that skips deep blocks when the saliency-weighted probe drift ( $\delta_{\mathrm{SWD}}$ ) is below a motion and step-adaptive threshold ( $\tau_{\mathrm{ATS}}(t)$ ) (computed by CFC+ATS), enabling cache hits. On a cache hit, OFA reuses computation by aligning cached residuals via optimal interpolation and optional motion-compensated warping to approximate ( $z_{t}^{(N)}$ ). On a cache miss, the model executes deep blocks, updates the residual cache in a ping-pong buffer, and continues the denoising loop for (t=1 $\ldots$ T).

Figure 3: Saliency overlay (Cosmos-Predict2.5, 2B). Channel-variance saliency from step 40 overlaid on four video frames. High saliency (yellow) marks structurally complex regions where caching errors are most visible; low saliency (purple) marks smooth areas where reuse is safe. SWD reweights drift using this signal to prioritize fidelity on detail-rich content.

Figure 4: Comparison of static vs. adaptive caching strategies . (a) A Fixed Threshold ( $\tau=0.12$ ) fails to accommodate the naturally increasing drift ( $\delta$ ) in later denoising steps, resulting in frequent recomputations and a low skip rate (36%). (b) Our Adaptive Threshold (ATS) dynamically scales $\tau_{ATS}$ with the expected drift. This allows ATS to capture significantly more cache hits (green bars) in later stages, nearly doubling the overall skip rate (68%) while maintaining generation quality.

	Domain Score								Quality Score									Avg. $\uparrow$	Lat. (s) $\downarrow$	Speedup $\uparrow$
Method	CS	AV	RO	IN	HU	PH	MI	Avg $\uparrow$	SC	BC	MS	AQ	IQ	OC	IS	IB	Avg $\uparrow$
Cosmos-Predict2.5 – 2B
Baseline	0.759	0.643	0.724	0.820	0.769	0.859	0.846	0.767	0.909	0.929	0.979	0.501	0.712	0.199	0.788	0.808	0.728	0.748	54.34	1.0 $\times$
DiCache	0.756	0.631	0.707	0.799	0.773	0.849	0.833	0.759	0.902	0.925	0.978	0.493	0.705	0.197	0.780	0.838	0.727	0.743	40.82	1.3 $\times$
FasterCache	0.675	0.553	0.549	0.691	0.652	0.719	0.745	0.629	0.849	0.909	0.970	0.405	0.594	0.176	0.709	0.796	0.676	0.652	34.51	1.6 $\times$
WorldCache	0.759	0.639	0.735	0.810	0.760	0.845	0.839	0.763	0.903	0.927	0.979	0.492	0.703	0.196	0.782	0.826	0.727	0.745	26.28	2.1 $\times$

Table 1: Text2World (T2W) generation results on PAI-Bench across two model scales. We evaluate four methods, Baseline (no acceleration), DiCache, FasterCache, and WorldCache (ours), on Cosmos-Predict2.5 at both 2B and 14B parameter scales. Domain Score aggregates seven semantic categories (CS: City Street, AV: Aerial View, RO: Road, IN: Indoor, HU: Human, PH: Physics, MI: Mixed), while Quality Score spans eight perceptual and fidelity dimensions (SC: Scene Consistency, BC: Background Consistency, MS: Motion Smoothness, AQ: Aesthetic Quality, IQ: Image Quality, OC: Object Consistency, IS: Imaging Subject, IB: Imaging Background). Avg. denotes the overall score averaged across both Domain and Quality metrics. Lat. reports wall-clock inference latency (seconds), and Speedup is relative to the unaccelerated Baseline.

\Delta

rows quantify the per-metric gain of WorldCache over DiCache, the strongest competing method. WorldCache achieves the best accuracy–efficiency trade-off, consistently outperforming DiCache while delivering up to

\mathbf{2.10\times}

speedup at 2B and

\mathbf{2.14\times}

at 14B with negligible quality degradation.

	Domain Score								Quality Score									Avg. $\uparrow$	Lat. (s) $\downarrow$	Speedup $\uparrow$
Method	CS	AV	RO	IN	HU	PH	MI	Avg $\uparrow$	SC	BC	MS	AQ	IQ	OC	IS	IB	Avg $\uparrow$
Cosmos-Predict2.5 – 2B
Baseline	0.919	0.694	0.811	0.877	0.840	0.909	0.886	0.845	0.896	0.929	0.982	0.505	0.674	0.212	0.936	0.952	0.761	0.803	55.04	1.0 $\times$
DiCache	0.899	0.697	0.791	0.876	0.828	0.887	0.909	0.835	0.885	0.923	0.980	0.492	0.660	0.212	0.927	0.940	0.752	0.794	39.68	1.4 $\times$
FasterCache	0.855	0.676	0.697	0.829	0.739	0.851	0.847	0.772	0.800	0.872	0.974	0.432	0.577	0.197	0.888	0.919	0.708	0.740	32.75	1.7 $\times$
WorldCache	0.912	0.708	0.796	0.876	0.833	0.893	0.890	0.840	0.892	0.926	0.982	0.496	0.661	0.212	0.931	0.948	0.756	0.798	24.48	2.3 $\times$

Table 2: Image2World (I2W) generation results on PAI-Bench across two model scales. We evaluate four methods: Baseline (no acceleration), DiCache, FasterCache, and WorldCache (ours) on Cosmos-Predict2.5 at both 2B and 14B parameter scales under the Image2World setting, where a conditioning image guides video world generation.

\Delta

rows quantify the per-metric gain of WorldCache over DiCache. Compared to the T2W setting, I2W scores are generally higher owing to the additional visual grounding provided by the conditioning image. WorldCache achieves the best accuracy–efficiency trade-off, consistently outperforming DiCache while delivering up to

\mathbf{2.3\times}

speedup at 2B and

\mathbf{2.18\times}

at 14B with negligible quality degradation across all domain and quality dimensions.

Task	Method	Domain Avg $\uparrow$	Quality Avg $\uparrow$	Overall $\uparrow$	Latency (s) $\downarrow$	Speedup $\uparrow$
WAN2.1 - 1.3B
T2W	Baseline	0.7862	0.7592	0.7727	120.04	1.00 $\times$
	DiCache	0.7841	0.7564	0.7703	61.57	1.96 $\times$
	WorldCache (Ours)	0.7853	0.7589	0.7721	50.84	2.36 $\times$
	$\Delta$ (WC $-$ DC)	+0.0012	+0.0025	+0.0018	$-$ 10.73	+0.40 $\times$
WAN2.1 - 14B

Table 3: WAN2.1 results on PAI-Bench. We report Domain Score (Avg), Quality Score (Avg), Overall Score (computed as the mean of Domain Avg and Quality Avg), wall-clock latency (seconds), and speedup relative to the unaccelerated baseline. Best accelerated results per task/column are in bold .

Figure 5: Qualitative Image2World comparison (PAI-Bench). Given the same conditioning image and text description, we show evenly-spaced frames from the generated rollout for (a) Baseline, (b) DiCache, and (c) WorldCache. DiCache exhibits temporal artifacts under dynamic scene evolution, e.g., object ghosting/deformation and inconsistent motion of vehicle and foreground structures (highlighted with red dashed boxes ). WorldCache preserves the scene layout and maintains more coherent object appearance and trajectories over time (highlighted with green dashed boxes ), while achieving substantial inference acceleration. Better viewed zoomed in.

Configuration	Domain $\uparrow$	Quality $\uparrow$	Overall $\uparrow$	Speedup $\uparrow$	Latency (s) $\downarrow$
Base	0.8447	0.7607	0.8027	1.00 $\times$	55
+ CFC	0.8457	0.7583	0.8020	1.52 $\times$	36
+ CFC + SWD	0.8414	0.7592	0.8003	1.67 $\times$	33
+ CFC + SWD + OFA	0.8468	0.7602	0.8035	1.49 $\times$	37
+ CFC + SWD + OFA + ATS ( WorldCache )	0.8395	0.7559	0.7977	2.3 $\times$	25

Table 4: Incremental ablation on Cosmos-Predict2.5-2B (I2W). Each row adds one module to the previous configuration. Legend: CFC = Causal Feature Caching (motion-adaptive thresholding); SWD = Saliency-Weighted Drift; OFA = Optimal Feature Approximation (Online System Identification); ATS = Adaptive Threshold Scheduling (dynamic threshold decay).

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Domain $\uparrow$	Quality $\uparrow$	Overall $\uparrow$	Lat. (s) $\downarrow$	Speedup $\uparrow$
Text2World (T2W) — Cosmos-Predict2.5-2B
Baseline	0.7670	0.7280	0.7475	54.34	1.0 $\times$
EasyCache	0.7641	0.7262	0.7451	41.41	1.3 $\times$
DiCache	0.7590	0.7272	0.7431	40.82	1.3 $\times$
TeaCache (Fast)	0.7616	0.7266	0.7448	41.07	1.4 $\times$
TeaCache (Slow)	0.7634	0.7274	0.7454	49.40	1.1 $\times$

Appendix

Table 5: PAI-Bench comparison with EasyCache and TeaCache (Cosmos-2B). We report Domain, Quality, Overall, latency (s), and speedup vs. baseline for both Image2World (I2W) and Text2World (T2W). Bold and underline denote the best and second-best among accelerated methods only (excluding Baseline) within each block.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Speedup $\uparrow$	Lat. (s) $\downarrow$
WAN2.1-14B (I2V)
Baseline	13.30	0.503	0.459	1.00 $\times$	391.9
DiCache	12.95	0.491	0.461	1.88 $\times$	208.6
WorldCache	13.19	0.498	0.460	2.30 $\times$	171.6
Cosmos-Predict-2.5-2B (I2V)
Baseline	12.87	0.455	0.518	1.00 $\times$	70.01

Appendix

Table 6: EgoDex-Eval results (I2V) across backbones. We report frame-level PSNR/SSIM/LPIPS against ground-truth videos, along with end-to-end latency and speedup relative to the unaccelerated baseline. Bold denotes the best value per block (including Baseline), and underline denotes the best accelerated result (DiCache/WorldCache).

Hyperparameter	Value	Domain $\uparrow$	Quality $\uparrow$	Overall $\uparrow$
CFC: $\alpha$	0.1	0.8389	0.7362	0.7876
	0.2	0.8457	0.7583	0.8020
	0.4	0.8417	0.7374	0.7896
	0.5	0.8362	0.7356	0.7859
SWD: $\beta_{s}$	0.05	0.8483	0.7335	0.7909
	0.12	0.8414	0.7592	0.8003

Appendix

Table 7: Hyperparameter sensitivity for skip decisions (stage-wise). Each block varies a single hyperparameter while keeping the rest fixed in the corresponding stage configuration: CFC is swept with only CFC enabled; SWD is swept with CFC fixed to

\alpha{=}2

; ATS is swept with CFC+SWD+OFA fixed to defaults. Results: Cosmos-Predict2.5–2B (I2W). Bold indicates the chosen default.

Hyperparameter	Value	Domain $\uparrow$	Quality $\uparrow$	Overall $\uparrow$
OFA operator	OSI only	0.8478	0.7360	0.7919
	Warp only	0.8402	0.7297	0.7850
	OSI + Warp	0.8468	0.7602	0.8035
Warp scale	0.2	0.8385	0.7309	0.7847
	0.5	0.8495	0.7420	0.7958
	1.0	0.8408	0.7357	0.7883

Appendix

Table 8: Hyperparameter sensitivity for cache-hit approximation (stage-wise). OFA operator and warp scale sweeps are run with CFC and SWD fixed to defaults. Results: Cosmos-Predict2.5–2B (I2W). Bold indicates the chosen default.

Figure 6: Effect of the denoising step budget. End-to-end latency and speedup of WorldCache vs. baselines when varying the number of denoising steps.

Figure 7: Qualitative Image2World comparison on PAI-Bench (Cosmos-2B). The conditioning image (top-left) and text description (top-right) specify a dashcam scene where the ego vehicle slows at a T-intersection as a woman and a child cross in front. We show evenly spaced frames from rollouts generated by (a) Baseline, (b) DiCache, and (c) WorldCache. DiCache exhibits temporal inconsistency on salient, moving entities, e.g., the pedestrians appearance and position become unstable and partially ghosted near the end of the rollout (red dashed box) . WorldCache maintains more coherent pedestrian identity and motion while preserving scene layout (green dashed box) , producing rollouts closer to the baseline under substantially lower inference latency. Better viewed zoomed in.

Figure 8: Qualitative Image2World comparison on PAI-Bench (Cosmos-14B). Given the conditioning image (top-left) and description (top-right), we generate a short kitchen rollout in which two people prepare food, where one spreads avocado on a nori sheet while the other holds a plate of vegetables. We visualize evenly spaced frames from (a) the unaccelerated baseline, (b) DiCache, and (c) WorldCache. While the baseline struggles with hand pose and object geometry, DiCache also exhibits noticeable temporal instability in the most salient, high-frequency regions, particularly around the hands and the plate: edges wobble, fine structures deform, and the appearance of the plate/hand region changes inconsistently over time ( red dashed boxes ). WorldCache largely avoids these artifacts, preserving sharper boundaries and more coherent local motion in the interaction region ( green dashed boxes ), yielding outputs that are closer to the baseline at substantially reduced latency. Best viewed with zoom.

Figure 9: Additional qualitative results. We show representative frames at increasing rollout times $(T_{0}\rightarrow T_{N})$ from PAI-Bench (top to bottom). A robot hand interacting with a refrigerator door, an egocentric manipulation scene near a tabletop, and an arm–human interaction setup. Across all examples, the generated videos maintain stable scene layout and consistent object/robot geometry as motion evolves over time, demonstrating that WorldCache preserves temporal continuity in contact-rich and interaction-heavy rollouts. Better viewed with zoom-in.

Figure 10: Additional qualitative results. We show evenly spaced frames from $(T_{0}\rightarrow T_{N})$ for three representative scenarios: robotic arm usage for kitchen activities, opening a refrigerator door, and interaction within a cluttered toolbox. The sequences highlight sustained temporal coherence where robot pose, object geometry, and scene layout remain stable as the action unfolds. Better viewed zoom-in.

Diffusion Model Caching

Diffusion model caching reuses features, activations, or attention states across denoising steps to reduce inference cost while preserving output quality.

See all 112 papers

WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching
Mar 6, 2026Weilun Feng, Guoxin Fan, Haotong Qin +10Diffusion Model Inference Acceleration Diffusion Model Caching
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences · ETH Zürich +4
SpectralCache: Accelerating Diffusion-Based World Models via Spectral Feature Caching
Oct 2, 2026Zhendong Mi, Pu Zhao, Ziyu Hu +4Diffusion Model Caching Diffusion Model Inference Acceleration
Stevens Institute of Technology · Northeastern University · Technische Universität Darmstadt
SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching
Feb 27, 2026Yasaman Haghighi, Alexandre AlahiVideo Diffusion Models Diffusion Model Caching
Ecole Polytechnique Fédérale de Lausanne (EPFL)

Mar 6, 2026J/K move · Enter open · S save

WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching

Weilun Feng, Guoxin Fan, Haotong Qin, Mingqiang Wu+9

State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences · ETH Zürich+4

Diffusion Model Inference Acceleration Diffusion Model Caching Interactive World Models World Model Learning

Open paper AskPDF arXiv

Diffusion-based world models have shown strong potential for unified world simulation, but the iterative denoising remains too costly for interactive use and long-horizon rollouts. While feature caching can accelerate inference without training, we find that policies designed for single-modal diffusion transfer poorly to world models due to two world-model-specific obstacles: token heterogeneity from multi-modal coupling and spatial variation, and non-uniform temporal dynamics where a small set of hard tokens drives error growth, making uniform skipping either unstable or overly conservative. We propose WorldCache, a caching framework tailored to diffusion world models. We introduce Curvature-guided Heterogeneous Token Prediction, which uses a physics-grounded curvature score to estimate token predictability and applies a Hermite-guided damped predictor for chaotic tokens with abrupt direction changes. We also design Chaotic-prioritized Adaptive Skipping, which accumulates a curvature-normalized, dimensionless drift signal and recomputes only when bottleneck tokens begin to drift. Experiments on diffusion world models show that WorldCache delivers up to \textbf{3.7

\times

} end-to-end speedups while maintaining 98% rollout quality, demonstrating the vast advantages and practicality of WorldCache in resource-constrained scenarios. Our code is released in https://github.com/FofGofx/WorldCache.

Abstract

Diffusion Model Caching

Explore similar work

WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching

SpectralCache: Accelerating Diffusion-Based World Models via Spectral Feature Caching

SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching