AI Paper Atlas

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

Tip-Adapter: Training-Free Adaption of CLIP for Few-Shot Classification

Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, Hongsheng Li

ECCV 2022 · Vision-language and multimodal › CLIP and zero-shot · approach: Fine-grained improvements to CLIP · 522 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 training-free key-value cache model to zero-shot CLIPZero-shot CLIP · ImageNet (16-shot)accuracy (%): 60.33 -> 62.03Improved
fine-tune the cached keys (Tip-Adapter-F)Tip-Adapter · ImageNet (16-shot)accuracy (%): 62.03 -> 65.51Improved
set residual ratio alpha to 1.0 instead of 0.0 (ablation)Tip-Adapter · ImageNet (16-shot)accuracy (%): 60.33 -> 62.03Improved
raise residual ratio alpha from 1.0 to 4.0 (ablation)Tip-Adapter · ImageNet (16-shot)accuracy (%): 62.03 -> 59.14Hurt
change sharpness beta from 1.5 to 5.5 (ablation)Tip-Adapter · ImageNet (16-shot)accuracy (%): 61.82 -> 62.03No clear effect
cache all 16 shots per class instead of none (ablation)Tip-Adapter · ImageNet (16-shot)accuracy (%): 60.33 -> 62.03Improved
increase training shots from 16 to 128 (cache fixed at 16 prototypes) (ablation)Tip-Adapter · ImageNet (16-shot)accuracy (%): 62.03 -> 63.15Improved
increase training shots from 16 to 128 (ablation)Tip-Adapter-F · ImageNet (16-shot)accuracy (%): 65.47 -> 69.74Improved
also fine-tune the cached values L_train (ablation)Tip-Adapter-F · ImageNet (16-shot)accuracy (%): 65.51 -> 60.90Hurt
fine-tune CLIP visual encoder instead of keys (ablation)Tip-Adapter · ImageNet (16-shot)accuracy (%): 62.03 -> 62.84Mixed
fine-tune CLIP textual encoder instead of keys (ablation)Tip-Adapter · ImageNet (16-shot)accuracy (%): 62.03 -> 63.15Mixed
jointly fine-tune both CLIP encoders (ablation)Tip-Adapter · ImageNet (16-shot)accuracy (%): 62.03 -> 51.22Hurt

Later papers that compared with this method

11 later accepted papers report TIP-Adapter as a baseline in their results tables; 11 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