data manifold
The data manifold is a geometric representation of the structure of the training data. It provides insights into the relationships and dimensions of the data, aiding in understanding the complexity and behavior of learned models.
- Convex Potential Mirror Langevin Algorithm for Efficient Sampling of Energy-Based Models
- Diffusion Models and the Manifold Hypothesis: Log-Domain Smoothing is Geometry Adaptive
- Generative Model Inversion Through the Lens of the Manifold Hypothesis
- Rectified CFG++ for Flow Based Models
- When and how can inexact generative models still sample from the data manifold?