Search papers, labs, and topics across Lattice.
The paper introduces a framework based on the string method to analyze the geometry of learned distributions in diffusion models by computing continuous paths between samples. This approach allows for interpolation between generative transport, minimum energy paths (MEPs), and finite-temperature string dynamics, enabling the exploration of the energy landscape without retraining. The key finding is that MEPs in image diffusion models produce unrealistic images, while principal curves yield realistic morphing sequences, and in protein structure prediction, the method computes physically plausible transition pathways.
Uncover hidden relationships in your diffusion model's learned distribution by tracing paths between samples with a new string method framework, revealing realistic morphing sequences and protein transition pathways.
Understanding the geometry of learned distributions is fundamental to improving and interpreting diffusion models, yet systematic tools for exploring their landscape remain limited. Standard latent-space interpolations fail to respect the structure of the learned distribution, often traversing low-density regions. We introduce a framework based on the string method that computes continuous paths between samples by evolving curves under the learned score function. Operating on pretrained models without retraining, our approach interpolates between three regimes: pure generative transport, which yields continuous sample paths; gradient-dominated dynamics, which recover minimum energy paths (MEPs); and finite-temperature string dynamics, which compute principal curves -- self-consistent paths that balance energy and entropy. We demonstrate that the choice of regime matters in practice. For image diffusion models, MEPs contain high-likelihood but unrealistic ''cartoon'' images, confirming prior observations that likelihood maxima appear unrealistic; principal curves instead yield realistic morphing sequences despite lower likelihood. For protein structure prediction, our method computes transition pathways between metastable conformers directly from models trained on static structures, yielding paths with physically plausible intermediates. Together, these results establish the string method as a principled tool for probing the modal structure of diffusion models -- identifying modes, characterizing barriers, and mapping connectivity in complex learned distributions.
