OriCache: Orientation-Guided Feature Caching for DiT Acceleration

Joonsik Nam, Hyunwoo Yu, Suk-Ju Kang

Keywords: Diffusion Transformers, Feature Caching, Cache Scheduling, Efficient Inference

TL;DR: OriCache shows that local update orientation can serve as a meaningful cue for training-free cache scheduling in DiT inference.

Abstract: Diffusion Transformers (DiTs) deliver high-quality generation but require repeated Transformer evaluations over many denoising steps, making inference expensive. Feature caching reduces this cost by reusing intermediate computations across denoising steps, where cache scheduling determines when cached computations can be reused or refreshed. Existing training-free scheduling methods have explored lightweight trajectory cues such as feature distance, update magnitude, and spectral variation. However, the orientation consistency of local updates remains underexplored, despite its potential to reflect the stability of local denoising trajectories. In this work, we propose OriCache, a training-free feature caching method that explicitly incorporates local update orientation into cache scheduling. OriCache computes a cache score from lightweight block inputs by combining update-scale variation with directional alignment between consecutive local updates. By jointly considering how much local updates change and whether their directions remain consistent, OriCache captures complementary aspects of denoising trajectory evolution for recomputation decisions. Experiments on FLUX.1-[dev] and Stable Diffusion 3.5-Large show that OriCache improves over representative static and adaptive caching baselines, and ablations confirm that combining magnitude and orientation yields more reliable cache decisions than either cue alone.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 33