OmniCache framework slashes diffusion model inference time by 35%

Researchers have developed OmniCache, a novel hierarchical caching framework designed to accelerate inference for high-resolution image and video diffusion models. By exploiting redundancy in intermediate diffusion features without altering model weights, OmniCache targets four types of redundancy: intra-frame, inter-frame, motion, and denoising-step. The framework employs a multidimensional feature reuse strategy through Token Cache, Frame Cache, Block Cache, and Layered Cache, which selects cacheable features based on similarity matching and skips redundant computations. This approach has demonstrated significant latency reductions of up to 35% for models like Stable Diffusion 3, while preserving visual fidelity and motion coherence in a training-free manner. AI

IMPACT This framework offers a significant speedup for diffusion models, potentially lowering inference costs and increasing accessibility for high-resolution image and video generation.

RANK_REASON This is a research paper detailing a new technical framework for diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Flux
• Frame Cache
• Latte
• Layered Cache
• OmniCache
• page cache
• Stable Diffusion 3
• SVD-XT
• Token Cache

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →