Newest papers

SpecKCast: Spectral Kalman Forecasting for Diffusion Acceleration

Javad Rajabi, Hakkı Can Karaimer, David B. Lindell, Babak Taati, Michael S. Brown · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26837 · cited by 0 works (OpenAlex) · open-access PDF

French

Les transformateurs de diffusion de vidéos (DiTs) bénéficient d'une image et d'une génération vidéo de haute fidélité, mais souffrent de coûts substantiels d'inférence en raison de l'évaluation des transformateurs répétée à travers de nombreuses étapes de dénouement. Le caching de la fonction réduit ce coût en réutilisant les fonctionnalités à travers les étapes de dénouement. Les méthodes de prédiction récentes étendent cette idée en extrapolant les fonctionnalités aux étapes échappées au lieu de les réutiliser. Cependant, les méthodes de prédiction existantes extrapolent les fonctionnalités qu'elles prédisent de manière uniforme, même si les composants de fréquence spatiale de ces fonctionnalités changent le plus à différents stades

Machine translation (Meta M2M100); the English original follows.

Abstract

Diffusion Transformers (DiTs) achieve high-fidelity image and video generation, but they suffer from substantial inference costs due to repeated transformer evaluations across many denoising steps. Feature caching reduces this cost by reusing features across denoising steps. Recent forecasting methods extend this idea by extrapolating features to skipped steps instead of reusing them. However, existing forecasting methods extrapolate the features they predict uniformly, even though the spatial-frequency components of these features change most at different stages of denoising. They also overlook that denoising progresses nonuniformly, so equal step intervals can correspond to very different amounts of change. These mismatches amplify forecast errors as more steps are skipped, limiting the achievable acceleration. We propose SpecKCast, a training-free method that forecasts the model output in the frequency domain. SpecKCast applies a 2D discrete cosine transform (DCT) to the output and tracks each coefficient with a lightweight Kalman filter along a band-specific time coordinate derived from the noise schedule. The filter tracks each coefficient's value and rate of change, and the transformer is evaluated only when the filter's forecast uncertainty exceeds a threshold. Experiments on FLUX.1-dev, Qwen-Image, Wan2.1-14B, and HunyuanVideo show that SpecKCast achieves a better trade-off between acceleration and fidelity than existing caching and forecasting methods, reaching up to 6x speedup with less than 0.05 GiB of additional memory.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Kalman filter, Computer science, Granularity, Inference, Algorithm, Noise reduction

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange