Diffusion Process with Implicit Latents via Energy Models
Keywords: Energy-based model, Image generation, Gradient Penalty, Diffusion Process
Abstract: We present a generative model based on an ordered sequence of latent variables for intermediate distributions between a given source and a desired target distribution. We construct the probabilistic transitions among the latent variables using energy models that are in the form of classifiers. In our work, the intermediate transitional distributions are implicitly defined by the energy models during training, where the statistical properties of the data distribution are naturally taken into account. This is in contrast to denoising diffusion probabilistic models (DDPMs) where they are explicitly defined by the predefined scheduling of a sequential noise degradation process. Over the course of training, our model is designed to optimally determine the intermediate distributions by Langevin dynamics driven by the energy model. In contrast, energy-based models (EBMs) typically require an additional generator since the intermediate distributions are not explicitly defined in the training procedure. We demonstrate the effectiveness and efficiency of the proposed algorithm in the context of image generation, achieving high fidelity results with less inference steps on a variety of datasets.
Supplementary Material: zip
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6443
Loading