Reproduce “Addressing degeneracies in latent interpolation for diffusion 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
Erik Landolsi, Fredrik Kahl
Researchers at Chalmers University of Technology develop a channel-wise mean adjustment technique for latent interpolation in diffusion models, addressing the quality degradation that occurs when interpolating between multiple latent vectors while maintaining image quality across both Stable Diffusion 1.5 and 3.5 models.
Key insights
- Latent vectors contain both deterministic bias and random noise components
- Using fixed normalization performs better than interpolating between input norms
- Quality degradation increases with the number of interpolated inputs
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.