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

The recent explosive interest on transformers has suggested their potentialto become powerful "universal" models for computer vision tasks, such asclassification, detection, and segmentation. While those attempts mainly studythe discriminative models, we explore transformers on some more notoriouslydifficult vision tasks, e.g., generative adversarial networks (GANs). Our goalis to conduct the first pilot study in building a GAN completely free ofconvolutions, using only pure transformer-based architectures. Our vanilla GANarchitecture, dubbed TransGAN, consists of a memory-friendly transformer-basedgenerator that progressively increases feature resolution, and correspondinglya multi-scale discriminator to capture simultaneously semantic contexts andlow-level textures. On top of them, we introduce the new module of gridself-attention for alleviating the memory bottleneck further, in order to scaleup TransGAN to high-resolution generation. We also develop a unique trainingrecipe including a series of techniques that can mitigate the traininginstability issues of TransGAN, such as data augmentation, modifiednormalization, and relative position encoding. Our best architecture achieveshighly competitive performance compared to current state-of-the-art GANs usingconvolutional backbones. Specifically, TransGAN sets new state-of-the-artinception score of 10.43 and FID of 18.28 on STL-10, outperforming StyleGAN-V2.When it comes to higher-resolution (e.g. 256 x 256) generation tasks, such ason CelebA-HQ and LSUN-Church, TransGAN continues to produce diverse visualexamples with high fidelity and impressive texture details. In addition, wedive deep into the transformer-based generation models to understand how theirbehaviors differ from convolutional ones, by visualizing training dynamics. Thecode is available at https://github.com/VITA-Group/TransGAN.

TransGAN: Two Pure Transformers Can Make One Strong GAN, and That Can Scale Up