AI Paper Atlas

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

Long-CLIP: Unlocking the Long-Text Capability of CLIP

Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Jiaqi Wang

ECCV 2024 · Vision-language and multimodal › CLIP and zero-shot · approach: CLIP training recipes and data curation · 265 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
add knowledge-preserved stretching of positional embedding instead of fixed-ratio interpolation (ablation)Long-CLIP · ImageNet zero-shottop-1 accuracy: 58.8 -> 66.8Improved
add primary component matching (ablation)Long-CLIP with KPS · CIFAR-100 zero-shottop-1 accuracy: 65.9 -> 69.3Improved
add primary component matching (ablation)Long-CLIP with KPS · Urban-200 text-to-imageR@1: 78.0 -> 79.0Improved
add KPS without PCM versus neither (ablation)Long-CLIP · COCO text-to-imageR@5: 43.4 -> 64.3Improved
naive fixed-ratio interpolation and long-text fine-tuning versus original CLIPCLIP ViT-B/16 · ImageNet zero-shottop-1 accuracy: 68.4 -> 55.1Hurt
use PCA-based primary component matching versus undistinguished image feature (ablation)Long-CLIP · Urban-200 text-to-imageR@1: 71.0 -> 79.0Improved
use PCM instead of mixed-length text (ablation)Long-CLIP · Urban-200 text-to-imageR@1: 67.5 -> 79.0Improved
use PCM instead of bounded text encoder (SmoothL1) (ablation)Long-CLIP · CIFAR-100 zero-shottop-1 accuracy: 67.9 -> 69.3Improved
Long-CLIP versus CLIP on zero-shot classificationCLIP ViT-L/14 · ImageNet zero-shottop-1 accuracy: 75.5 -> 73.5Hurt
Long-CLIP replaces CLIP on short-caption retrievalCLIP ViT-B/16 · COCO 5k text-to-imageR@1: 32.7 -> 40.4Improved
apply KPS and PCM to DeCLIP ViT-B/32DeCLIP · Urban long-caption retrievalT2I/I2T R@1: 28.0/25.5 -> 54.0/51.0Improved

Later papers that compared with this method

17 later accepted papers report Long-CLIP as a baseline in their results tables; 16 of them report a better number on most of the results they share with it.

9 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