English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Tools annual

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

Despite the tantalizing success in a broad of vision tasks, transformers havenot yet demonstrated on-par ability as ConvNets in high-resolution imagegenerative modeling. In this paper, we seek to explore using pure transformersto build a generative adversarial network for high-resolution image synthesis.To this end, we believe that local attention is crucial to strike the balancebetween computational efficiency and modeling capacity. Hence, the proposedgenerator adopts Swin transformer in a style-based architecture. To achieve alarger receptive field, we propose double attention which simultaneouslyleverages the context of the local and the shifted windows, leading to improvedgeneration quality. Moreover, we show that offering the knowledge of theabsolute position that has been lost in window-based transformers greatlybenefits the generation quality. The proposed StyleSwin is scalable to highresolutions, with both the coarse geometry and fine structures benefit from thestrong expressivity of transformers. However, blocking artifacts occur duringhigh-resolution synthesis because performing the local attention in ablock-wise manner may break the spatial coherency. To solve this, weempirically investigate various solutions, among which we find that employing awavelet discriminator to examine the spectral discrepancy effectivelysuppresses the artifacts. Extensive experiments show the superiority over priortransformer-based GANs, especially on high resolutions, e.g., 1024x1024. TheStyleSwin, without complex training strategies, excels over StyleGAN onCelebA-HQ 1024, and achieves on-par performance on FFHQ-1024, proving thepromise of using transformers for high-resolution image generation. The codeand models will be available at https://github.com/microsoft/StyleSwin.

StyleSwin: Transformer-based GAN for High-resolution Image Generation