Post-hoc Discriminator Guidance For Data-efficient Image Generation Via Annealing Langevin Dynamics
Keywords: post-hoc discriminator guidance, GANs, data-efficient image generation, annealing Langevin dynamics
TL;DR: Post-hoc discriminator guidance helps data-efficient image generation for better sampling quality via annealing Langevin dynamics.
Abstract: The proposed method, post-hoc discriminator guidance (PDG) aims to take an alternate route for Nash non-equilibrium issue in GANs' training. This method introduces an additional discriminator that gives explicit supervision with regard to gradient of density ratio $\nabla_{x} \log_{}{\frac{p_{r}(x)}{p_{f}(x)}}$ between real and fake probability density function, steering the sample path towards more realistic regions in a post-hoc way. We train the discriminator after adversarial optimization, making post-hoc discriminator training stable and fast to converge. In generation process, annealing Langevin dynamics sampling with density ratio score reduces the Kullback-Leibler divergence between the true and generated samples. Given an optimal discriminator, the method can improve the sampling quality of various off-the-shelf models on the web without retraining required. Extensive experiments validate the advancements and effectiveness of PDG on content-varying data-limited datasets.
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: 2903
Loading