BRACE: Taming Sharp Irregularities via Barycentric Rational Forecasting for Fast Diffusion Transformers Inference

Jinlong Yang1, Jinke WU1, Lizilin1 Yao Zhou1, *
1Sichuan University
*Corresponding author
ACM Multimedia 2026
Paper Code arXiv
BRACE Teaser

BRACE addresses caching bottlenecks in generative visual models, ensuring high-fidelity generation while drastically reducing computational overhead.

Abstract

Diffusion Transformers (DiT) have demonstrated exceptional capabilities in visual generation, but their inherently sequential denoising process leads to substantial inference latency. Previous acceleration strategies, such as caching intermediate feature maps, often struggle with sharp irregularities across timesteps, leading to error accumulation and degraded generation fidelity.

In this work, we propose BRACE, a novel mathematical framework leveraging Barycentric Rational Forecasting to accurately extrapolate and tame these sharp irregularities. Extensive experiments across various generative tasks demonstrate that BRACE consistently outperforms existing methods in the latency-quality trade-off, enabling fast inference without structural modifications.

Method

BRACE sequential inference algorithm BRACE system-level update pipeline and forecast micro-architecture
BRACE method overview. Sequential inference alternates exact computation with barycentric rational forecasting through a local sliding window, domain mapping, and a lightweight forecast micro-architecture.
PC1 feature trajectory, cosine similarity, L1 error, and MSE error comparisons across diffusion sampling steps
Empirical evidence. Barycentric rational forecasting closely follows the ground-truth feature trajectory while maintaining higher cosine similarity and lower L1 and MSE errors across sampling steps.

Qualitative Results