Efficient Controllable Generation with Guarantee

Xiaojiang Yang, Jiacheng Sun, Xing Zhang, Zhenguo Li, Junchi Yan

Keywords: controllable generation, variational autoencoder, identifiability

Abstract: Generative models have achieved great success in image synthesis, and controllability of the generative process is a key requirement for their successful adoption in real-world applications. Most existing methods for controllable generation lack theoretical guarantees and are time-consuming, which weakens their reliability and applicability. In this paper, we propose an identifiability theorem to provide a guarantee of controllability. This theorem ensures that semantic attributes can be disentangled and hence independently controlled by orthogonalization in latent space in a supervised manner. Based on the theoretical analysis, we propose a general method for controllable generation, which can be integrated with most latent-variable generative models. We further propose to plug it into a pre-trained NVAE. Such a scheme significantly reduces the cost of time and has better consistency in image editing due to the merits of NVAE. Experiments show that our method is comparable with the state-of-the-art methods in attribute-conditional generation and image editing, and has advantages in efficiency and consistency.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: Generative models

TL;DR: We propose a general method with theoretical guarantee and integrate it with NVAE for controllable generation.