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.

```python
def causal_feature_caching(feature, motion_magnitude, base_threshold=0.02):
    """
    Motion-aware skip decision: reuse features when motion is small.
    During fast dynamics (high motion), enforce stricter thresholds.
    """
    # Estimate motion from optical flow or frame differencing
    motion_scale = compute_motion_magnitude(feature)

    # Scale threshold inversely with motion: high motion → stricter threshold
    adaptive_threshold = base_threshold / (1.0 + motion_scale)

    # Skip reuse decision
    should_reuse = motion_scale < adaptive_threshold
    return feature if should_reuse else recompute(feature)
```

#### 2. Saliency-Weighted Drift (SWD)
Prioritizes perceptually important regions so errors on salient objects matter more than background noise.

```python
def saliency_weighted_drift(feature, cached_feature, saliency_map):
    """
    Compute drift penalty weighted by perceptual importance.
    Salient regions incur high cost for mismatch; background errors are tolerated.
    """
    drift = feature - cached_feature
    weighted_loss = (drift ** 2) * saliency_map

    return weighted_loss.mean()
```

#### 3. Optimal Feature Approximation (OFA)
Uses least-squares blending and motion-compensated warping instead of verbatim copying.

```python
def optimal_feature_approximation(feature, cached_feature, motion_field):
    """
    Blend cached and current features via least-squares optimization.
    Compensate for motion-induced misalignment via optical flow warping.
    """
    # Motion-compensated warp of cached feature
    warped_cached = warp_features(cached_feature, motion_field)

    # Least-squares blend: find optimal α
    alpha = solve_lstsq(feature, warped_cached)
    approximated = alpha * feature + (1 - alpha) * warped_cached

    return approximated
```

#### 4. Adaptive Threshold Scheduling (ATS)
Relaxes thresholds during late denoising refinement while maintaining strict control during structure formation.

```python
def adaptive_threshold_scheduling(timestep, total_steps, base_threshold=0.02):
    """
    Threshold schedule: strict early (structure formation), relaxed late (refinement).
    Early: timestep/total_steps >> 0 (high noise)
    Late: timestep/total_steps << 0 (low noise, small updates)
    """
    progress = 1.0 - (timestep / total_steps)  # 0 at start, 1 at end

    # Early steps strict, late steps aggressive
    scheduled_threshold = base_threshold * (1.0 + 2.0 * progress)

    return scheduled_threshold
```

## Practical Integration

### When to Apply
- Generative video or image diffusion models
- Real-time or latency-sensitive inference scenarios
- Multi-frame prediction or frame interpolation tasks
- Hardware with bandwidth constraints (inference on edge devices)

### Integration Steps
1. **Instrument codebase**: Extract motion magnitude and saliency per timestep
2. **Profile baseline**: Measure per-step latency and memory bandwidth on your hardware
3. **Deploy CFC module**: Start with motion-aware skip decisions; tune base_threshold for your model
4. **Add SWD**: Compute saliency map (via gradient or CLIP) and weight drift penalties
5. **Implement OFA**: Replace feature copying with least-squares blending + motion warping
6. **Schedule thresholds**: Adjust aggressiveness across denoising steps
7. **Validate quality**: Ensure metrics match 99.4%+ retention; adjust parameters if lower

### Tuning Parameters
- `base_threshold`: Start at 0.02; lower for stricter reuse (quality-focused), higher for aggressive speedup
- `motion_scale_factor`: Control how motion magnitude scales threshold; typically 1.0–2.0
- `saliency_weighting`: Emphasize important regions; adjust based on task (faces more salient than backgrounds)
- `late_phase_aggressiveness`: Late denoising steps tolerate more approximation; typically 2×–4× base threshold