AI Paper Atlas

Research map / Vision-language and multimodal / CLIP and zero-shot

CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts

Yichao Cai, Yuhang Liu, Zhen Zhang, Javen Qinfeng Shi

ECCV 2024 · Vision-language and multimodal › CLIP and zero-shot · approach: CLIP training recipes and data curation · 14 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
apply contrastive learning with augmented prompts (CLAP) to CLIP representationsCLIP ViT-B/16 · Zero-shot ZS(C) overallavg top-1 accuracy (%): 77.4 -> 79.9Improved
apply CLAP to CLIP under noised promptCLIP ViT-B/16 · Zero-shot ZS(NC) overallavg top-1 accuracy (%): 70.4 -> 76.1Improved
use image augmentation (Im.Aug) instead of prompt augmentation (ablation)CLIP ViT-B/16 · Zero-shot ZS(C) overallavg top-1 accuracy (%): 77.4 -> 76.1Hurt
use image augmentation (Im.Aug) on OfficeHome (ablation)CLIP ViT-B/16 · OfficeHome ZS(C)top-1 accuracy (%): 79.8 -> 77.0Mixed
switch from image augmentation to prompt augmentation (ablation)CLAP · Prompt variance overallrange R: 1.9 -> 1.1Improved
apply CLAPCLIP ViT-B/16 · Few-shot 1-shot linear probe PACStop-1 accuracy gain: +10%Improved
apply CLAP to 1-shot classifier under PGD-20CLIP ViT-B/16 · Adversarial robustness 1-shot PGD-20accuracy gain: +11.7%Improved
apply CLAP to zero-shot classifier under CW-20CLIP ViT-B/16 · Adversarial robustness ZS(C) CW-20accuracy gain: +1.1%No clear effect
apply Im.Aug to zero-shot classifier under CW-20 (ablation)CLIP ViT-B/16 · Adversarial robustness ZS(C) CW-20accuracy gain: 0%No clear effect
vary latent dimension, temperature and inference weight (ablation)CLAP · VLCS zero-shotavg top-1 accuracyNo clear effect
use simpler training prompt forms rather than LLM or random-style prompts (ablation)CLAP · VLCS zero-shotavg top-1 accuracyImproved
train disentangled network on text prompts instead of synthetic images (ablation)CLAP · Training time DomainNettraining time: 3.3 hours -> about 47 minutesImproved

Later papers that compared with this method

11 later accepted papers report CLAP as a baseline in their results tables; 7 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