worldcache-video-model-caching

name	worldcache-video-model-caching
title	WorldCache: Efficient Inference in Diffusion World Models via Content-Aware Caching
version	0.0.3
engine	skillxiv-v0.0.3-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2603.22286
keywords	["Video Diffusion","Caching","Inference Optimization","World Models","Perceptual Quality"]
description	Replace zero-order hold assumptions with perception-constrained approximation via four-module caching system: motion-aware skip thresholds, saliency-weighted drift, least-squares blending, and adaptive scheduling. Achieves 2.1–2.3× speedup at 2B scale with 99.4–99.6% quality retention on Cosmos-Predict video generation; apply when optimizing diffusion world model inference under real-time constraints.

Resource Definition

Four-module content-aware caching system for video diffusion model inference:

1. Causal Feature Caching (CFC): Motion-aware skip decisions
2. Saliency-Weighted Drift (SWD): Perceptual importance prioritization
3. Optimal Feature Approximation (OFA): Intelligent feature blending
4. Adaptive Threshold Scheduling (ATS): Time-dependent reuse aggressiveness

Efficiency Dimension

Primary optimization target: Wall-clock latency of generative step in video/image diffusion models.

Secondary targets: Memory bandwidth, GPU utilization during cached-feature reuse.

Trade-off axis: Reuse aggressiveness (throughput gain) vs. perceptual quality preservation.

Performance-Efficiency Tradeoff

Speedup vs. Quality

The framework achieves its dramatic speedup by investing in decision quality early, then "spending" the quality margin aggressively in later denoising phases where updates become small refinements.

Cosmos-Predict2.5 Results:

Scale	Task	Speedup	Quality Retention	vs. DiCache	vs. FasterCache
2B	Text-to-World	2.1×	99.6%	+0.8× faster	+0.4× faster
14B	Text-to-World	2.14×	~99.6%	+0.8× faster	+0.4× faster
2B	Image-to-World	2.3×	~99.6%	Superior quality	Superior quality

Competing Methods:

• DiCache: Only 1.3–1.4× speedup with visible quality degradation
• FasterCache: 1.6–1.7× speedup with noticeable artifacts

Overall Quality: PAI-Bench measurements confirm ~99.4% quality retention across diverse evaluation scenarios.

Four-Module Architecture

1. Causal Feature Caching (CFC)

Adapts skip thresholds based on input motion magnitude, preventing aggressive reuse during fast dynamics.

def causal_feature_caching(feature, motion_magnitude, base_threshold=0.02):
    
    
    motion_scale = compute_motion_magnitude(feature)

    
    adaptive_threshold = base_threshold / ( + motion_scale)

    
    should_reuse = motion_scale < adaptive_threshold
     feature  should_reuse  recompute(feature)