Reproduce “PixelRush: Ultra-Fast, Training-Free High-Resolution Image Generation via One-step Diffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hong-Phuc Lai, Phong Nguyen, Anh Tran
PixelRush makes high-resolution images by upscaling a coarse SDXL image, splitting it into overlapping patches, partially inverting them with DDIM, and applying brief SDXL-Turbo refinement. It preserves structure while adding detail and uses blending and noise injection to reduce artifacts. On the 4K task, it reports FID 54.67 in 20 seconds versus FreeScale’s 58.28 in 323 seconds under different upsampling configurations.
Key insights
- Executing a full reverse diffusion process to perturb the latent to full Gaussian noise is computationally redundant for refinement tasks, as partial inversion can achieve similar results more efficiently.
- Few-step diffusion models can be successfully adapted for high-resolution refinement by preserving structural information through partial inversion and employing specialized blending and noise injection techniques.
- Traditional patch blending methods introduce severe artifacts in few-step contexts due to large, sharp updates within patches, necessitating a Gaussian filter-inspired feathering blend for seamless transitions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.