ProCache: Constraint-Aware Feature Caching with Selective Computation for Diffusion Transformer Acceleration

Fanpu Cao, Yaofo Chen, Zeng You, Wei Luo, Cen Chen

Abstract

Diffusion Transformers (DiTs) have set new benchmarks in high-fidelity image and video generation but suffer from significant computational overhead. ProCache is a training-free acceleration framework that replaces traditional, uniform feature caching with model-tailored, non-uniform schedules. By combining an offline search for optimal caching patterns with selective partial updates of critical tokens and layers, ProCache achieves up to a 2.9x speedup on state-of-the-art models like DiT-XL/2 and FLUX.1 with negligible loss in image quality.

Introduction

Diffusion Transformers (DiTs) are the architectural backbone of leading generative models such as Sora and Stable Diffusion 3. Despite their performance, their massive parameter counts make inference slow and resource-intensive. A common acceleration strategy is feature caching—reusing computed data from one denoising step in subsequent steps to skip redundant calculations.

However, current caching methods suffer from two primary flaws. First, they typically employ uniform schedules (e.g., recomputing every $N$ steps), which fails to account for the fact that DiT features are stable early in the denoising process but change rapidly toward the end. Second, naive reuse leads to error accumulation, particularly in deeper network layers, which degrades image quality at high acceleration ratios.

Methodology: The ProCache Framework

ProCache introduces a two-pronged strategy to optimize caching without requiring model retraining: a constraint-aware pattern search and a selective computation mechanism.

Instead of a human-defined uniform schedule, ProCache utilizes an offline search to identify a non-uniform binary sequence (compute vs. cache). This search is guided by three logical constraints:

  • Budget: Adhering to a total allowed compute limit.
  • Monotonicity: Ensuring intervals between computations shorten as the denoising process becomes more dynamic.
  • Boundedness: Preventing excessively long caching gaps that could lead to model divergence.

2. Selective Computation

During steps designated for caching, ProCache performs a "lightweight refresh" rather than a total skip. It selectively recomputes only the deepest blocks—where errors tend to amplify—and only for the most important tokens, identified by the magnitude of their attention outputs. This partial update adds minimal latency (~3%) while significantly stabilizing the generation process.

Overview of the ProCache framework, illustrating the constraint-aware pattern search and the selective computation mechanism.

Empirical Motivation: Temporal Dynamics

The core insight behind ProCache's non-uniform schedule stems from an analysis of temporal redundancy. By plotting the L1 error between consecutive denoising steps, the authors observed an exponential-like curve. Features remain remarkably stable for the first 30 steps but evolve violently in the final 10. This empirical evidence suggests that "front-loading" the cache and "back-loading" computation is the most efficient path to acceleration.

L1 error between consecutive denoising steps, showing the exponential increase in feature dynamics toward the end of the process.

Key Results and Evaluation

ProCache was evaluated across several benchmarks, including ImageNet (DiT-XL/2), MS-COCO (PixArt-α), and PartiPrompts (FLUX.1). It consistently outperformed prior state-of-the-art training-free methods like FORA and Δ-DiT.

  • Efficiency: ProCache achieved a 2.90× speedup on DiT-XL/2 and 1.96× on PixArt-α.
  • Quality: On ImageNet, ProCache maintained an FID of 2.96, significantly outperforming uniform caching methods (FID 3.88) at comparable speeds.
  • FLUX Acceleration: On the highly optimized FLUX.1-dev model, ProCache delivered a 1.54× speedup. Interestingly, it also improved the "Image Reward" score compared to the baseline, suggesting the non-uniform schedule helps the model focus on critical structural details.

Conclusion and Future Work

ProCache demonstrates that intelligent, non-uniform caching is a powerful tool for accelerating DiTs without the need for costly retraining. While currently focused on image generation, the framework is naturally suited for video DiTs (like Sora), which exhibit even higher temporal redundancy. Future research may explore more sophisticated token importance metrics, such as those involving cross-attention or gradient information, to further refine the selective computation process.

Summary updated

BibTeX
@inproceedings{cao2026procache,
  author = {Fanpu Cao and Yaofo Chen and Zeng You and Wei Luo and Cen Chen},
  title = {{ProCache: Constraint-Aware Feature Caching with Selective Computation for Diffusion Transformer Acceleration}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i24.39069},
}
Figure