Latent Express Traversal for Manifold Denoising

We introduce Latent Express Traversal (LET), a memory- and compute-efficient denoising-by-traversal framework. By exploiting the low dimensional subspace-manifold hierarchy of the representation, the traversal is performed directly in the latent space. Furthermore, to improve the robustness, we endow our optimizer with funnels that serve as effective pathways to escape local minima. By moving along the geometric manifold in latent space with this augmented optimizer, our method substantially reduces the overall memory overhead, while enhancing denoising accuracy. On gravitational-wave denoising, LET achieves a substantially better accuracy--memory trade-off than ambient-space manifold traversal, while funnels further improve accuracy with only a modest increase in model size.