AI Paper Atlas

Research map / Image generation / Editing and style transfer

Diagonal Attention and Style-Based GAN for Content-Style Disentanglement in Image Generation and Translation

Gihyun Kwon, Jong Chul Ye

ICCV 2021 · Image generation › Editing and style transfer · approach: Image-to-image translation · 56 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use diagonal attention (DAT) content control instead of SNI input tensorStyleGAN generator · CelebA-HQ 256 (with per-pixel noise)W PPL: 58.21 -> 48.12Improved
use DAT content control instead of SNI input tensorStyleGAN generator · AFHQ 256 (with per-pixel noise)W PPL: 65.22 -> 63.44Improved
use DAT instead of SNI without per-pixel noiseStyleGAN generator · AFHQ 256 (w/o per-pixel noise)W PPL: 127.99 -> 73.51Improved
add DS loss to SNI baselineSNI · AFHQ 256 (with per-pixel noise)W PPL: 65.22 -> 69.70Hurt
use DAT instead of SNI (image quality)StyleGAN generator · CelebA-HQ 256 (with per-pixel noise)FID: 10.79 -> 10.90No clear effect
set DS loss lambda to 0.3 rather than 0.2 (ablation)Diagonal GAN · CelebA-HQ 256 (with per-pixel noise)W PPL: 49.83 -> 48.12Improved
set DS loss lambda to 0.3 rather than 0.4 (ablation)Diagonal GAN · AFHQ 256 (with per-pixel noise)W PPL: 80.41 -> 63.44Improved
set DS loss lambda to 0.3 rather than 0.5 (ablation)Diagonal GAN · CelebA-HQ 256 (with per-pixel noise)W PPL: 68.75 -> 48.12Improved
use single-layer MLP rather than 2-layer MLP for attention mapping (ablation)Diagonal GAN · CelebA-HQ 256 (with per-pixel noise)W PPL: 53.99 -> 48.12Improved
use single-layer MLP rather than CNN for attention mapping (ablation)Diagonal GAN · CelebA-HQ 256 (with per-pixel noise)W PPL: 83.37 -> 48.12Improved
extend DAT layers from up to 32x32 to up to 256x256 (ablation)Diagonal GAN · AFHQ 256 (with per-pixel noise)W PPL: 118.21 -> 63.44Improved
extend DAT layers from up to 64x64 to up to 256x256 (ablation)Diagonal GAN · CelebA-HQ 256 (with per-pixel noise)W PPL: 61.98 -> 48.12Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy