Reproduce “ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models” 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

Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, Sungroh Yoon

Researchers from Seoul National University and Samsung SDS developed ILVR, a learning-free method for conditioning pre-trained unconditional Denoising Diffusion Probabilistic Models. This technique enables a single unconditional DDPM to perform diverse controlled generation tasks such as image translation and editing, while maintaining high visual quality.

Key insights

  • Low-pass filtering operations can effectively capture coarse semantic information from reference images.
  • Injecting low-frequency information from a reference image at each step of the reverse diffusion process enables controlled generation without retraining.
  • The degree of semantic control and output diversity can be intuitively adjusted by varying the downsampling factor and the conditioning range.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.