AI Paper Atlas

Research map / Image generation / Editing and style transfer

AdaAttN: Revisit Attention Mechanism in Arbitrary Neural Style Transfer

Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, Errui Ding

ICCV 2021 · Image generation › Editing and style transfer · approach: Arbitrary neural style transfer · 464 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 cosine-similarity attention instead of softmax (ablation)AdaAttN · Video style transfer (mean of 20 styles)optical flow error (x1e-2): 5.52 -> 4.09Improved
add cross-image similarity loss L_is (ablation)AdaAttN · Video style transfer (mean of 20 styles)optical flow error (x1e-2): 5.52 -> 5.51No clear effect
add cross-image similarity loss L_is on top of cosine attention (ablation)AdaAttN + Cos · Video style transfer (mean of 20 styles)optical flow error (x1e-2): 4.09 -> 3.91Improved
replace local feature loss with L2 content loss (ablation)AdaAttN · Image style transfervisual qualityHurt
add local feature loss instead of content loss (ablation)AdaAttN · Image style transfervisual qualityImproved
add global style loss to local feature loss (ablation)AdaAttN · Image style transfercolor saturationImproved
add shallow features to Q and K (ablation)AdaAttN · Image style transfervisual qualityImproved
use AdaAttN instead of SANet for videoAdaAttN vs SANet · Video style transfer (mean of 20 styles)optical flow error (x1e-2): 7.76 -> 5.52Improved
add cosine attention and L_is to beat Linear and MCCNetAdaAttN + Cos + L_is · Video style transfer (mean of 20 styles)optical flow error (x1e-2): 4.42 -> 3.91Improved
use AdaAttN instead of SANet for speedAdaAttN vs SANet · Inference time, 512x512seconds per image: 0.081 -> 0.112Hurt

Later papers that compared with this method

11 later accepted papers report AdaAttN as a baseline in their results tables; 9 of them report a better number on most of the results they share with it.

3 more not listed.

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