article

A new DBMI estimator uses discrete diffusion bridges to accurately estimate mutual information for discrete data, performing well on low-dimensional and image tasks.

This paper introduces IDLM, a method that accelerates discrete Diffusion Language Models by adapting inverse distillation techniques, achieving 4x-64x faster inference while maintaining generative quality through theoretically-validated unique solutions and stable gradient relaxations.

The work unifies significant parts of the generative modeling landscape. It shows that CFM and (forward-only) IFM are dual perspectives of the same approach, while positioning general IFM as a broader, more expressive paradigm that encompasses and extends beyond CFM. This bridge deepens theoretical understanding and opens avenues for practical innovation in both lines of research.

Current EOT models are slow or limited. We introduce VarEOT, a faster, simpler method that avoids simulation. It matches or beats other methods in tasks like image translation.

We propose Electric Current Discrete Data Generation (ECD2G), a novel method for discrete data generation that uses a neural network to learn electric currents in a circuit, which represent a provable probability flow for transferring a source distribution to a target.

This paper introduces Inverse Poisson Flow Matching (IPFM), a new distillation method that significantly accelerates electrostatic generative models like PFGM by learning a generator to match the teacher’s electrostatic field, enabling high-quality sample generation in few steps.

We introduce a novel distillation technique for ResShift, one of the best diffusion models for real-world image super-resolution.

We propose a novel light learning algorithm for semi-supervised domain translation which can utilize both paired and unpaired data.

We reveal several limitations of existing solvers for the continuous Gromov-Wasserstein problem and pave the way to get rid of them by proposing a novel solver.