AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

CLIP-Adapter: Better Vision-Language Models with Feature Adapters

Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, Yu Qiao

IJCV 2023 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 1,625 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 residual feature adapter on frozen CLIP image featuresZero-shot CLIP RN50 · ImageNet (16-shot)accuracy: 55.41% -> 61.33%Improved
replace CoOp prompt tuning with CLIP-AdapterCoOp RN50 · ImageNet (16-shot)accuracy: 60.46% -> 61.33%Improved
use CLIP-Adapter instead of CoOp for training timeCoOp RN50 · ImageNet (16-shot)train time: 14h 40min -> 50minImproved
add residual connection (alpha 0.6 instead of 1.0) (ablation)CLIP-Adapter · DTD (16-shot)accuracy: 63.79 -> 66.06Improved
add residual connection (alpha 0.2 instead of 1.0) (ablation)CLIP-Adapter · ImageNet (16-shot)accuracy: 59.05 -> 61.33Improved
insert adapter only at end of encoder instead of all layers (ablation)CLIP-Adapter ViT-B/16 · ImageNet (16-shot)accuracy: 67.67 -> 70.88Improved
insert adapter at layer 10 instead of layer 12 (ablation)CLIP-Adapter ViT-B/16 · ImageNet (16-shot)accuracy: 70.88 -> 71.85Mixed
set bottleneck dimension to D/4 instead of D (ablation)CLIP-Adapter · ImageNet (16-shot)accuracy: 59.78 -> 61.33Improved
shrink bottleneck dimension from D/4 to D/32 (ablation)CLIP-Adapter · ImageNet (16-shot)accuracy: 61.33 -> 59.45Hurt
also finetune CLIP visual encoder (ablation)CLIP-Adapter · ImageNet (16-shot)accuracy: 61.33% -> 60.07%Hurt
also finetune both CLIP encoders (ablation)CLIP-Adapter · ImageNet (16-shot)accuracy: 61.33% -> 52.78%Hurt
use ensemble of 7 hard prompts (ablation)CLIP-Adapter · ImageNet (16-shot)accuracy: 61.33 -> 61.68Improved

Later papers that compared with this method

8 later accepted papers report CLIP-Adapter as a baseline in their results tables; 8 of them report a better number on most of the results they share with it.

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