StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation

Abstract

We explore and analyze the latent style space of Style-GAN2, a state-of-the-art architecture for image generation, using models pretrained on several different datasets. We first show that StyleSpace, the space of channel-wise style parameters, is significantly more disentangled than the other intermediate latent spaces explored by previous works. Next, we describe a method for discovering a large collection of style channels, each of which is shown to control a distinct visual attribute in a highly localized and disentangled manner. Third, we propose a simple method for identifying style channels that control a specific attribute, using a pretrained classifier or a small number of example images. Manipulation of visual attributes via these StyleSpace controls is shown to be better disentangled than via those proposed in previous works. To show this, we make use of a newly proposed Attribute Dependency metric. Finally, we demonstrate the applicability of StyleSpace controls to the manipulation of real images. Our findings pave the way to semantically meaningful and well-disentangled image manipulations via simple and intuitive interfaces.

Attribute-specific Control Units in StyleGAN for Fine-grained Image Manipulation

“…It suggests that our method can achieve fine-grained controls on local semantic regions of generated images. Note that, our proposed method can perform various local attribute editing tasks, which is much more than previous methods [10,25,26,33]. Then we further visualize the control units for several attribute manipulations in Figure 4.…”

“…[26,28] further show that meaningful directions can be computed in closed form directly from the generator's weights without any form of training or optimization. Recent methods [7,19,33] reveal that the walks in S space facilitates the spatial disentanglement in the spatial dimension. [7] accomplished the local semantically-aware edits by transferring modulation style between source and target image.…”

“…[7] accomplished the local semantically-aware edits by transferring modulation style between source and target image. StyleSpace [33] shows that finegrained controls can be obtained by only modifying the value of one single detected channel. However, modifying only one channel may also affect multiple attributes of the image or only have a slight effect on the synthesized image's target attribute.…”

“…This means, while the target attribute of the generated images is successfully manipulated using these methods, irrelevant regions and attributes in the image are usually also changed. To address this issue, very recent works [19,33] further introduce the S space spanned by the modulation styles. StyleSpace [33] measures and compares the disentanglement and completeness of these spaces using the metrics proposed by [8], and finds that the feature map's modulation styles at different scales are rich in semantic information.…”

“…To address this issue, very recent works [19,33] further introduce the S space spanned by the modulation styles. StyleSpace [33] measures and compares the disentanglement and completeness of these spaces using the metrics proposed by [8], and finds that the feature map's modulation styles at different scales are rich in semantic information. Spatial disentangled manipulation can be achieved by modifying only a single channel of the manipulation style.…”