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

2025-12-19

AI Summary

To address error accumulation and limited efficiency in feature caching for real-time deployment of Diffusion Transformers (DiTs)—caused by fixed-interval caching and full-feature reuse—this paper proposes a training-free dynamic feature caching framework. Leveraging the time-varying evolution of DiT features, the framework introduces a temporal-aware, non-uniform caching scheduler that enables selective computation at both depth and token levels. It further pioneers a synergistic mechanism combining constraint-aware cache pattern search with lightweight cache state prediction to support adaptive, online caching decisions. Evaluated on PixArt-α and DiT models, the method achieves 1.96× and 2.90× inference speedup, respectively, while incurring less than 0.5% degradation in FID and CLIP scores—substantially outperforming existing approaches.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Feature Construction/Reformulation
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Diffusion Transformers (DiTs) have achieved state-of-the-art performance in generative modeling, yet their high computational cost hinders real-time deployment. While feature caching offers a promising training-free acceleration solution by exploiting temporal redundancy, existing methods suffer from two key limitations: (1) uniform caching intervals fail to align with the non-uniform temporal dynamics of DiT, and (2) naive feature reuse with excessively large caching intervals can lead to severe error accumulation. In this work, we analyze the evolution of DiT features during denoising and reveal that both feature changes and error propagation are highly time- and depth-varying. Motivated by this, we propose ProCache, a training-free dynamic feature caching framework that addresses these issues via two core components: (i) a constraint-aware caching pattern search module that generates non-uniform activation schedules through offline constrained sampling, tailored to the model's temporal characteristics; and (ii) a selective computation module that selectively computes within deep blocks and high-importance tokens for cached segments to mitigate error accumulation with minimal overhead. Extensive experiments on PixArt-alpha and DiT demonstrate that ProCache achieves up to 1.96x and 2.90x acceleration with negligible quality degradation, significantly outperforming prior caching-based methods.

Problem

• Accelerates Diffusion Transformers for real-time use
• Optimizes non-uniform caching to match DiT dynamics
• Reduces error accumulation in feature reuse

Innovation

• Non-uniform caching schedules via constrained sampling
• Selective computation in deep blocks and high-importance tokens
• Training-free dynamic feature caching for diffusion transformers