Reproduce “DiffRGD: An Inference-Time Diffusion Guidance Through Riemannian Gradient Descent” on OpenResearch

About this paper

Jia-Wei Liao, Li-Xuan Peng, Mei-Heng Yueh, Min Sun, Cheng-Fu Chou, Jun-Cheng Chen

DiffRGD introduces an inference-time diffusion guidance method that explicitly preserves the latent Gaussian distribution during sampling by framing each step as a constrained optimization problem on a dynamically defined spherical manifold. This approach, solved via Riemannian Gradient Descent, yields higher quality images with fewer artifacts and improved adherence to conditions across various image restoration and conditional generation tasks.

Key insights

• Isotropic Gaussian latent variables in DDIMs can be geometrically decomposed and understood as residing on a dynamically defined spherical manifold at each timestep.
• Explicitly preserving the latent Gaussian distribution through a distribution-aware geometric constraint is crucial for mitigating sample quality degradation during inference-time guidance.
• Riemannian Gradient Descent offers a theoretically sound and effective mechanism for performing gradient-based guidance while maintaining the intrinsic geometry of the latent space.

Summary by alphaXiv.