CutDiffusion: A Simple, Fast, Cheap, and Strong Diffusion Extrapolation Method

TL;DR

CutDiffusion is a tuning-free, efficient method that enhances high-resolution diffusion extrapolation by splitting the process into structure denoising and detail refinement, achieving faster, cheaper, and stronger results.

Contribution

It introduces a simple, tuning-free approach that improves high-resolution diffusion extrapolation by dividing the process into two focused phases, reducing costs and increasing performance.

Findings

Single-step inference speeds up high-resolution diffusion.

Fewer patches needed for structure denoising reduces GPU costs.

Strong generation quality from focused detail refinement.

Abstract

Transforming large pre-trained low-resolution diffusion models to cater to higher-resolution demands, i.e., diffusion extrapolation, significantly improves diffusion adaptability. We propose tuning-free CutDiffusion, aimed at simplifying and accelerating the diffusion extrapolation process, making it more affordable and improving performance. CutDiffusion abides by the existing patch-wise extrapolation but cuts a standard patch diffusion process into an initial phase focused on comprehensive structure denoising and a subsequent phase dedicated to specific detail refinement. Comprehensive experiments highlight the numerous almighty advantages of CutDiffusion: (1) simple method construction that enables a concise higher-resolution diffusion process without third-party engagement; (2) fast inference speed achieved through a single-step higher-resolution diffusion process, and fewer inference patches required; (3) cheap GPU cost resulting from patch-wise inference and fewer patches during the comprehensive structure denoising; (4) strong generation performance, stemming from the emphasis on specific detail refinement.