CacheScope speeds up diffusion models by up to 3.57x and cuts LPIPS by 70%

The authors of an arXiv preprint have proposed CacheScope, an add-on for feature caching in diffusion transformers. The method speeds up inference by up to 3.57 times and lowers median LPIPS to 70.1% relative to the strongest probe-based baseline. The tests cover three models: FLUX.1-dev for images, and Wan2.1-T2V-1.3B and Open-Sora 1.2 for video. No fine-tuning, gradients or auxiliary networks are required.

Where feature caching breaks down

The logic of training-free acceleration is simple. On selected denoising steps, the model does not compute all blocks. Instead, it substitutes an approximation from a cache built on earlier steps. Researchers have so far put most of their effort into reconstructing the skipped features more accurately, designing ever more complex reconstructors.

The CacheScope authors argue that this approach has hit a ceiling. Once a stable first-order reconstructor is in place, further sophistication adds little. What matters far more for quality is something else: on which steps the model computes in full and on which it takes the cache. By their data, where the exact steps are placed matters considerably more than how refined the approximation is.

A risk signal instead of raw variation

That leaves the question of how to tell whether the cache is dangerous at a given step. Existing probe-based methods look at how much the output of a shallow probe has changed. The authors point out a problem: this absolute variation depends on two factors at once. The first is how far the input trajectory has shifted. The second is how strongly the network reacts to that shift. The raw value therefore does not directly reflect caching risk.

Their solution is normalization. The probe's output variation is divided by the corresponding input shift, which yields a directional gain per unit of trajectory movement. That gain serves as the indicator.

Based on it, CacheScope finds, online, the steps where reusing the cache is high-risk and switches to full computation on them. The cache reconstructor stays the same. The method works without gradients, additional training, auxiliary networks or offline schedule optimization for a specific prompt.

What the numbers claim

• Speedup: up to 3.57x relative to full denoising.
• Quality: median LPIPS lower, by up to 70.1%, than the strongest probe-based baseline.
• Models: FLUX.1-dev, Wan2.1-T2V-1.3B, Open-Sora 1.2.
• Requirements: no training, gradients, auxiliary networks or prompt-specific schedules.
• Cache reconstructor: unchanged; the method is added on top of an existing one.

The "up to" wording matters. The maximum speedup and the maximum LPIPS improvement most likely come from different models and settings. The abstract does not say what hardware was used or how the numbers change under more conservative settings. This is a first-version preprint, with no peer review and no independent reproductions.

What to check in practice

For teams that already use feature caching on FLUX or video models, one architectural detail stands out: CacheScope does not replace the reconstructor. It decides when the reconstructor must not be used. In theory, that lets you plug it into an existing pipeline without rewriting the cache and without a separate per-prompt calibration stage.

There are three things to check before adopting it. First, compare the method with your current cache on real prompts, not only by LPIPS: the metric is useful but does not replace visual evaluation. Second, measure the overhead of the probe and the online decisions on your own hardware, because that overhead determines the net gain. Third, look at behavior across different trajectory lengths and step counts. For video, where artifacts accumulate across frames, this is especially critical.

If the claimed numbers hold up, the result would be a cheap way to shift the speed-quality balance without fine-tuning the models.