Reproduce “NoiseDiffusion: Correcting Noise for Image Interpolation with Diffusion Models beyond Spherical Linear Interpolation” 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
PengFei Zheng, Yonggang Zhang, Zhen Fang, Tongliang Liu, Defu Lian, Bo Han
NoiseDiffusion presents a method for image interpolation using diffusion models that addresses the artifact generation often seen when interpolating natural images. The approach corrects non-standard noise encoding from natural images, leading to interpolated outputs with fewer artifacts and enhanced preservation of original image features compared to prior techniques.
Key insights
- Artifacts in natural image interpolation with diffusion models stem from the encoded noise not conforming to the model's expected distribution, leading to components outside the denoising threshold.
- Maintaining latent variables on a hypersphere, consistent with high-dimensional Gaussian properties, is crucial for the diffusion model's denoising process and for artifact mitigation.
- A balance must be struck between correcting invalid noise, introducing appropriate noise for interpolation, and preserving original image information to achieve high-quality, artifact-free results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.