Javad Rajabi, Hakki Can Karaimer, David B. Lindell, Babak Taati, Michael S. Brown · 2026 · arXiv
DOI: 10.48550/arXiv.2610.26837 · open-access PDF
Os Transformadores de Difusão (DiTs) relatam alta fidelidade de imagem e geração de vídeo, mas sofrem de custos substanciais de inferência devido a avaliações de transformadores repetidas em muitas etapas de denoissing. O caching de recursos reduz esse custo por reutilizar recursos em etapas de denoissing. Métodos de previsão recentes estendem essa ideia extrapolando recursos para etapas esquecidas em vez de reutilizá-los. No entanto, os métodos de previsão existentes extrapolam os recursos que eles preveem uniformemente, mesmo que os componentes de frequência espacial desses recursos mudem mais em diferentes estágios de denoissing.
Machine translation (Meta M2M100); the English original follows.
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 overlo
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.26837, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange