Reproduce “Smooth Diffusion: Crafting Smooth Latent Spaces in 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
Jiayi Guo, Xingqian Xu, Yifan Pu, Zanlin Ni, Chaofei Wang, Manushree Vasu, Shiji Song, Gao Huang and 1 more
Researchers from SHI Labs at Georgia Tech & UIUC, Picsart AI Research, and Tsinghua University developed Smooth Diffusion, a method that regularizes diffusion models to create smooth latent spaces. This approach yields significantly smoother interpolations and more accurate image inversion and editing capabilities, achieving an ISTD of 16.54 compared to Stable Diffusion's 38.63 while maintaining high generation quality.
Key insights
- Latent space smoothness in diffusion models can be enforced effectively through a step-wise regularization, making global smoothness achievable during the training process.
- Minimizing the proposed regularization loss encourages the diffusion system's Jacobian to be orthogonal, ensuring a consistent and predictable mapping from latent space perturbations to output image changes.
- Achieving a smoother latent space is complementary to, and can even slightly improve, the overall image generation quality (FID, CLIP Score) of diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.