AI Paper Atlas

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

VILA: Learning Image Aesthetics From User Comments With Vision-Language Pretraining

Junjie Ke, Keren Ye, Jiahui Yu, Yonghui Wu, Peyman Milanfar, Feng Yang

CVPR 2023 · Vision-language and multimodal › CLIP and zero-shot · approach: Language models for zero-shot reasoning · 75 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
Pretrain on aesthetic image-comment pairs in addition to general image-text data (ablation)VILA-P · AVAzero-shot SRCC: 0.228 -> 0.657Improved
Add both general and aesthetic pretraining before the adapter (ablation)VILA-R · AVASRCC: 0.566 -> 0.774Improved
Use a rank-based target instead of L2 score regression (ablation)VILA-R · AVASRCC: 0.757 -> 0.774Improved
Use a rank-based target instead of EMD loss (ablation)VILA-R · AVASRCC: 0.759 -> 0.774Improved
Use the frozen text prompt embedding as the ranking anchor (ablation)VILA-R · AVASRCC: 0.763 -> 0.774Improved
Add a residual connection to the learnable image projection (ablation)VILA-R · AVASRCC: 0.766 -> 0.774Improved
Finetune the image encoder rather than keep it frozen (ablation)VILA-R · AVASRCC: 0.774 -> 0.780Improved
Finetune the image encoder (ablation)VILA-P · AVA-StylemAP: 69.0% -> 26.3%Hurt
Set the contrastive-to-generative loss ratio to 1:2 (ablation)VILA-P · AVAzero-shot SRCC: 0.645 -> 0.657Improved
Set the ranking margin to 0.1 (ablation)VILA-R · AVASRCC: 0.769 -> 0.774Improved
Randomly sample a comment for each image during aesthetic pretraining (ablation)VILA-P · AVAzero-shot PLCC: 0.596 -> 0.663Improved
Use ensemble prompts instead of a single prompt (ablation)VILA-P · AVAzero-shot SRCC: 0.605 -> 0.657Improved

Later papers that compared with this method

14 later accepted papers report VILA1.5-13B as a baseline in their results tables; 12 of them report a better number on most of the results they share with it.

6 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