Relational Feature Caching for Accelerating Diffusion Transformers
Byunggwan Son, Jeimin Jeon, Jeongwoo Choi, Bumsub Ham
TL;DR
This paper introduces relational feature caching (RFC), a novel framework that improves diffusion transformer acceleration by leveraging input-output relationships for more accurate feature prediction, reducing redundant computations and outperforming prior methods.
Contribution
The paper proposes relational feature caching (RFC) with relational feature estimation and cache scheduling to enhance feature prediction accuracy in diffusion transformers.
Findings
RFC significantly outperforms prior caching approaches.
Relational feature estimation improves prediction accuracy.
Adaptive cache scheduling reduces unnecessary full computations.
Abstract
Feature caching approaches accelerate diffusion transformers (DiTs) by storing the output features of computationally expensive modules at certain timesteps, and exploiting them for subsequent steps to reduce redundant computations. Recent forecasting-based caching approaches employ temporal extrapolation techniques to approximate the output features with cached ones. Although effective, relying exclusively on temporal extrapolation still suffers from significant prediction errors, leading to performance degradation. Through a detailed analysis, we find that 1) these errors stem from the irregular magnitude of changes in the output features, and 2) an input feature of a module is strongly correlated with the corresponding output. Based on this, we propose relational feature caching (RFC), a novel framework that leverages the input-output relationship to enhance the accuracy of the…
Peer Reviews
Decision·ICLR 2026 Poster
This work introduces novel components—Relational Feature Estimation (RFE) and Relational Cache Scheduling (RCS)—that have not been explored in prior work. It utilizes input–output relationships, enabling a more accurate prediction of output features. The paper clearly articulates the motivation, challenges, and contributions, and provides a general framework that could inspire further research for DiTs.
Clarity of Mathematical Formulation (RELATIONAL FEATURE CACHING section): The presentation of equations in the RELATIONAL FEATURE CACHING section lacks clarity, making it difficult for readers to follow the mathematical foundations of the approach. For instance, the connection between the two components, RFE and RCS, is not clear at the beginning, and the logical flow from one equation to the next is not always well-motivated. Actionable suggestion: Including intuitive descriptions or intermed
This paper studies an important topic of feature caching, which is critical for optimizing the efficiency of diffusion transformers. The identification of input–output correlation as a predictor of output changes is insightful and well supported by the empirical analysis. The proposed RFC showcase on multiple metrics and settings, demonstrating a comparable to superior performance. The authors also present qualitative results, which are promising.
The organization of the manuscript needs improvement. There is overlapping and duplicate content across the first three sections. It may be better to defer the detailed discussion of related work to a later section and to reorganize Section 2 and Section 3.1. While the empirical correlation between inputs and outputs is demonstrated, the paper offers little formal analysis explaining why this correlation should hold across arbitrary architectures or datasets.
- Clear motivation with strong empirical evidence showing the limitation of purely temporal forecasting. - Simple yet effective design—RFE and RCS are well-justified and complementary. - Comprehensive experiments across multiple diffusion models and tasks.
- In Table 2, TaylorSeer achieves the highest CLIP scores; the paper should discuss why RFC does not consistently outperform it on semantic alignment metrics. - The paper could analyze RFC’s applicability to U-Net–based diffusion models to better demonstrate generality and architectural adaptability.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Memory and Neural Computing · Parallel Computing and Optimization Techniques · Evolutionary Algorithms and Applications