WorldCache: Content-Aware Caching for Accelerated Video World Models

Accepted at ECCV 2026 · European Conference on Computer Vision

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

WorldCache is a training-free, plug-and-play inference acceleration framework for diffusion-based video world models. It achieves up to 3.0× speedup while strictly maintaining temporal coherence and visual fidelity.

---

🎬 Qualitative Preview

High-fidelity video generation on Cosmos-Predict 2.5 (14B) with upto 3.0x speedup.

---

---

📖 Abstract

Video World Models increasingly rely on large-scale diffusion transformers to simulate complex spatial dynamics. However, the high computational cost of autoregressive generation remains a significant bottleneck. WorldCache overcomes this by identifying temporal and spatial redundancies in the denoising process.

Unlike naive caching which causes "motion drift," WorldCache uses a suite of content-aware modules like Causal Feature Caching (CFC), Saliency-Weighted Drift (SWD), Optimal Feature Approximation (OFA), and Adaptive Threshold Scheduling (ATS) to predict skipped computation rather than blindly copying it. Our method generalizes across leading architectures like NVIDIA Cosmos, WAN2.1, and DreamDojo.

---

✨ Key Components

WorldCache is driven by four key technical ideologies:

Module	Icon	Description
Causal Feature Caching (CFC)	⚡	Dynamically scales caching tolerance based on early layer motion velocity.
Saliency-Weighted Drift (SWD)	🎯	Penalizes caching errors in perceptually critical high-frequency regions.
Optimal Feature Approx. (OFA)	🌊	Interpolates skipped cache states using trajectory matching and optical flow.
Adaptive Threshold Scheduling (ATS)	📈	Exponentially relaxes caching constraints in later denoising stages.

---

🔬 Method Overview

WorldCache treats caching like a localized prediction. It controls the pace with causal tracking while interpolating the next state.

🛠️ Technical Highlights

• Drift Probing: Uses the first $K$ blocks of the transformer as a lightweight proxy for global drift.
• Motion-Adaptive Thresholds: Uses $\alpha$-scaled motion signals to prevent "ghosting" artifacts in high-dynamics scenes.
• Saliency Mapping: Weights L1 drift by spatial saliency (channel-wise variance) to preserve fine details.

---

⚙️ Installation & Setup

For detailed system requirements, environment setup (Virtual Env/Docker), and checkpoint downloading instructions, please refer to our:

👉 Detailed Setup Guide

⚡ Quick Summary (Conda + UV)

# 1. Create and activate conda environment
conda create -n worldcache python=3.10 -y
conda activate worldcache

# 2. Sync dependencies with UV
curl -LsSf https://astral.sh/uv/install.sh | sh
cd Models/Cosmos-Predict2.5/
uv sync --extra=cu128 --active --inexact

# 3. Basic inference (from root)
python Models/Cosmos-Predict2.5/examples/inference.py --model 2B/post-trained --worldcache_enabled [options]

---

🚀 Quick Start

To generate high-quality video with WorldCache acceleration:

# From the root of the repository - Text2World
CUDA_VISIBLE_DEVICES=0 python Models/Cosmos-Predict2.5/examples/inference.py \
  -i Models/Cosmos-Predict2.5/path/to/prompt.json \
  -o outputs/worldcache_output \
  --inference-type=text2world \
  --model 2B/post-trained \
  --disable-guardrails \
  --worldcache_enabled \
  --worldcache_motion_sensitivity 2.0 \
  --worldcache_flow_enabled \
  --worldcache_flow_scale 2.0 \
  --worldcache_osi_enabled \
  --worldcache_saliency_enabled \
  --worldcache_saliency_weight 1.0 \
  --worldcache_dynamic_decay

# From the root of the repository - Image2World
CUDA_VISIBLE_DEVICES=0 python Models/Cosmos-Predict2.5/examples/inference.py \
  -i Models/Cosmos-Predict2.5/path/to/prompt.json \
  -o outputs/worldcache_output \
  --inference-type=image2world \
  --model 2B/post-trained \
  --disable-guardrails \
  --worldcache_enabled \
  --worldcache_motion_sensitivity 2.0 \
  --worldcache_flow_enabled \
  --worldcache_flow_scale 2.0 \
  --worldcache_osi_enabled \
  --worldcache_saliency_enabled \
  --worldcache_saliency_weight 1.0 \
  --worldcache_dynamic_decay

---

📊 Quantitative Results

WorldCache establishes a new state-of-the-art for training-free diffusion acceleration, maintaining near-baseline quality while significantly reducing latency.

---

🌐 Model & Benchmark Coverage

Model Family	Scales	Architecture	PAI-Bench	EgoDex-Eval
Cosmos-Predict 2.5	2B, 14B	DiT	✅	✅
WAN2.1	1.3B, 14B	DiT	✅	✅
DreamDojo	2B	DiT	—	✅

---

1. PAI-Bench: Physical Reasoning Benchmarks

Across two major architectures (Cosmos and WAN), WorldCache consistently delivers >2× speedup with <1% drop in overall physical reasoning scores.

Table 1: PAI-Bench Text-to-World (T2W) Results

Model	Method	Domain Avg	Quality Avg	Overall	Latency (s)	Speedup
Cosmos 2B	Baseline	0.767	0.728	0.748	54.34	1.00×
	DiCache	0.759	0.727	0.743	40.82	1.3×
	WorldCache	0.763	0.727	0.745	26.28	2.1×
Cosmos 14B	Baseline	0.792	0.746	0.769	216.25	1.00×
	DiCache	0.792	0.745	0.768	148.36	1.4×
	WorldCache	0.795	0.746	0.771	98.61	2.14×

Table 2: PAI-Bench Image-to-World (I2W) Results

Model	Method	Domain Avg	Quality Avg	Overall	Latency (s)	Speedup
Cosmos 2B	Baseline	0.845	0.761	0.803	55.04	1.00×
	DiCache	0.835	0.752	0.794	39.68	1.40×
	WorldCache	0.840	0.756	0.798	24.48	2.30×
Cosmos 14B	Baseline	0.860	0.769	0.814	210.07	1.00×
	DiCache	0.855	0.767	0.811	146.04	1.40×
	WorldCache	0.859	0.768	0.813	99.25	2.18×

---

2. Architecture Transfer: WAN2.1

WorldCache is backbone-agnostic. On the latest WAN2.1 architecture, it achieves superior speed-quality tradeoffs compared to DiCache.

Table 3: WAN2.1 Transfer Results

Backbone	Method	Overall	Latency (s)	Speedup
T2W 1.3B	Baseline	0.7727	120.04	1.00×
	DiCache	0.7703	61.57	1.96×
	WorldCache	0.7721	50.84	2.36×
I2W 14B	Baseline	0.7384	475.60	1.00×
	DiCache	0.7311	291.91	1.53×
	WorldCache	0.7388	206.73	2.31×

---

3. EgoDex-Eval: Robotics Performance

In egocentric robotics tasks requiring high spatial precision, WorldCache maintains frame-level fidelity (PSNR/SSIM) while enabling real-time-friendly inference.

Table 4: EgoDex-Eval (Robotics Evaluation)

Model	Method	PSNR	SSIM	LPIPS	Latency (s)	Speedup
WAN2.1-14B	Baseline	13.30	0.503	0.459	391.90	1.00×
	DiCache	12.95	0.491	0.461	208.60	1.88×
	WorldCache	13.19	0.498	0.460	171.60	2.30×
Cosmos-2.5-2B	Baseline	12.87	0.455	0.518	70.01	1.00×
	DiCache	12.63	0.445	0.531	51.97	1.34×
	WorldCache	12.82	0.466	0.518	43.24	1.62×
DreamDojo-2B	Baseline	23.63	0.775	0.226	19.73	1.00×
	DiCache	20.41	0.734	0.252	12.46	1.58×
	WorldCache	23.69	0.737	0.251	10.36	1.90×

---

📈 Denoising Step Budget Scaling

WorldCache scales effectively with the denoising step budget. For longer generation trajectories (more steps), the efficiency gains increase as the underlying motion manifold stabilizes.

Efficiency scaling: WorldCache achieves up to 3.1× speedup as denoising steps increase, while maintaining superior quality over DiCache.

---

🖼️ Qualitative Gallery

WorldCache maintains flawless temporal coherence across diverse domains, from urban traffic to precision robotics.

Main comparison: WorldCache stays closer to baseline rollout in dynamic and interaction-heavy regions.

---

🙏 Acknowledgements

We acknowledge the following works that inspired this project:

• Cosmos-Predict 2.5 — NVIDIA's world foundation model platform.
• WAN2.1 — Open suite of video foundation models.
• DreamDojo — Generalist robot world model from the NVIDIA GEAR Team.
• DiCache — DiCache: Let Diffusion Model Determine Its Own Cache, ICLR 2026.

---

📝 Citation

@inproceedings{nawaz2026worldcache,
  title         = {WorldCache: Content-Aware Caching for Accelerated Video World Models},
  author        = {Umair Nawaz and Ahmed Heakl and Ufaq Khan and Abdelrahman Shaker and Salman Khan and Fahad Shahbaz Khan},
  booktitle     = {European Conference on Computer Vision (ECCV)},
  eprint        = {2603.22286},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  note          = {Accepted},
  year          = {2026}
}

📄 License

This project inherits the Apache 2.0 License from the NVIDIA Cosmos-Predict2 codebase.