![[2602.19506] Relational Feature Caching for Accelerating Diffusion Transformers](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.19506] Relational Feature Caching for Accelerating Diffusion Transformers
Summary
This paper introduces Relational Feature Caching (RFC) to enhance the efficiency of diffusion transformers by improving feature prediction accuracy and reducing computational redundancy.
Why It Matters
As machine learning models, particularly diffusion transformers, become more complex, optimizing their performance is crucial. RFC addresses significant prediction errors in existing caching methods, potentially leading to more efficient AI applications in various fields, including computer vision and generative AI.
Key Takeaways
- RFC improves the accuracy of feature predictions by leveraging input-output relationships.
- The framework introduces relational feature estimation (RFE) to better predict output changes.
- Relational cache scheduling (RCS) minimizes full computations based on expected prediction errors.
- Extensive experiments show RFC outperforms previous caching methods significantly.
- This approach can lead to more efficient AI models, enhancing their practical applications.
Computer Science > Computer Vision and Pattern Recognition arXiv:2602.19506 (cs) [Submitted on 23 Feb 2026] Title:Relational Feature Caching for Accelerating Diffusion Transformers Authors:Byunggwan Son, Jeimin Jeon, Jeongwoo Choi, Bumsub Ham View a PDF of the paper titled Relational Feature Caching for Accelerating Diffusion Transformers, by Byunggwan Son and 3 other authors View PDF HTML (experimental) 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 feature prediction. Specifically, we introduce relational feature estimation (RFE) to estimate the magnitude of changes in the output features from the inputs, enabling more accurate fea...
