Gaussian diffusion models have become a practical backbone for generative modeling. Denoising Lévy Probabilistic Models (DLPM) extend this idea by injecting heavy-tailed perturbations, but they rely on a fixed tail index throughout the entire denoising process.

The new arXiv preprint (2610.19692v1) introduces an annealed version of this approach. As the title indicates, the tail index is annealed — varied over the course of generation — while the reverse dynamics are replaced by a Gaussian surrogate. The abstract in the source is truncated after "We introduce Annealed…", so the precise annealing schedule and experimental results are not available from this source alone.

Because there is only a single source, there are no independent results to compare; the claims are the authors' own. The significance, as far as the abstract shows, is a more flexible middle ground between Gaussian diffusion and heavy-tailed Lévy perturbations.