Can We Generate Realistic Hands Using Only Convolution?

Mehran Hosseini, Peyman Hosseini

Keywords: Computer Vision, Generative Models, AI Generated Hands, Convolutional Neural Networks, Generative Adversarial Networks, Variational AutoEncoders

TL;DR: Towards alleviating abnormal details in images synthesized by generative models via augmenting convolutional layers with coordinate information.

Abstract: Despite their extensive use in generating hyperrealistic images, image generative models are prone to generating abnormal details and malformed features. One of the most prominent examples of this phenomenon is synthesizing contorted and mutated hands and fingers. Convolution serves as the backbone of many state-of-the-art image generative models, all of which are subject to the aforementioned phenomenon. We investigate whether adding a single channel, comprising horizontal and vertical coordinate information, to the input channels of convolution layers can alleviate this issue. We show that the answer is “yes”! We demonstrate this, in a GPU-poor setup, on two families of generative models, Generative Adversarial Networks (GANs) and Variational AutoEncoders (VAEs) trained on the Hand Gesture dataset. The hand images generated by models employing our method surpass those of models using simple convolution by a significant margin. We further validate the results for generating human faces using models trained on the CelebA-HQ dataset, demonstrating our models consistently yield superior images compared to those generated using simple convolution.

Primary Area: representation learning for computer vision, audio, language, and other modalities

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/2024/AuthorGuide.

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: 7877