PTL-Diffusion: Manifold-Aware Diffusion with Periodic Terminal Laws

Researchers propose PTL-Diffusion, a novel diffusion model framework that replaces single Gaussian terminal distributions with periodic families of Gaussian laws to better capture manifold structure in data. The approach embeds phase information directly into forward process dynamics rather than only in the denoising network, showing improved performance on point-cloud and facial datasets compared to standard DDPM baselines.

PTL-Diffusion addresses a fundamental limitation in current diffusion models: their reliance on unstructured terminal reference distributions that poorly represent data concentrated on low-dimensional manifolds. Standard diffusion models collapse all data toward a single Gaussian at the end of the forward process, forcing the reverse model to recover complex geometric structure from minimal information. This new framework introduces periodic terminal laws—multiple Gaussian distributions that vary cyclically through the forward process—embedding phase structure directly into the noising dynamics rather than treating it as a downstream conditioning signal.

The theoretical contribution remains grounded in standard diffusion principles: the authors derive closed-form forward marginals and explicit Gaussian reverse posteriors for their periodically forced Ornstein-Uhlenbeck process, enabling standard noise-prediction training without fundamental architectural changes. This design choice maintains compatibility with existing training procedures while introducing structured inductive biases about manifold geometry. The invariant-average regularization term couples phase-conditioned reverse dynamics through the averaged periodic reference law, providing a principled way to leverage the periodic structure.

Experimental validation demonstrates meaningful improvements on torus, cylinder, and Olivetti face datasets, with measurable reductions in phase-conditioned errors, feature-space covariance mismatches, and nearest-neighbor manifold distances. These gains suggest that explicitly incorporating geometric priors into terminal distributions can enhance generation quality for structured data. The work opens avenues for more sophisticated phase constructions and scaling to higher-dimensional datasets, potentially benefiting applications requiring faithful geometric representation such as 3D object generation, molecular simulation, or medical imaging.

• →PTL-Diffusion replaces single Gaussian terminal distributions with periodic families to better represent manifold-structured data.
• →Phase information is embedded directly into forward process dynamics rather than only conditioning the denoising network.
• →Closed-form solutions enable standard training procedures without architectural redesign.
• →Experiments show measurable improvements in manifold-level distributional matching and geometric accuracy.
• →The approach suggests structured terminal reference laws as a promising research direction for future diffusion model development.