AI Paper Atlas

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

b-CLIP: Text-Conditioned Contrastive Learning for Multi-Granular Vision-Language Alignment

Fatimah Zohra, Chen Zhao, Hani Itani, Bernard Ghanem

CVPR 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: CLIP training recipes and data curation

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 text-conditioned multi-granular alignment (b-CAL, K=6, CE) versus plain global-loss fine-tuning (ablation)CLIP ViT-B/16 · FG-OVD Hardaccuracy: 22.0 -> 29.2Improved
add multiple sentence positives with plain global loss (no text-conditioning) (ablation)CLIP fine-tuning · FG-OVD Hardaccuracy: 22.0 -> 19.7Hurt
use beta > 0 instead of strict self-matching (CE) (ablation)b-CLIP CE K=36 · FG-OVD Hardaccuracy: 3.6 -> 30.9Improved
use beta > 0 instead of strict self-matching (BCE) (ablation)b-CLIP BCE K=36 · FG-OVD Hardaccuracy: 14.8 -> 20.2Improved
raise beta from 0.5 to 1 (CE) (ablation)b-CLIP CE K=36 · FG-OVD Hardaccuracy: 30.9 -> 30.8No clear effect
increase hierarchy size K from 6 to 36 (CE) (ablation)b-CLIP CE · FG-OVD Hardaccuracy: 29.2 -> 30.9Improved
increase hierarchy size K from 6 to 36 (BCE) (ablation)b-CLIP BCE · FG-OVD Medium/Easyaccuracy: Easy 45.3 -> 34.2Hurt
use BCE instead of CE loss (ablation)b-CLIP · Urban-1k T2IR@1: 89.0 -> 91.8Improved
use BCE instead of CE loss (ablation)b-CLIP · FG-OVD Hardaccuracy: 30.9 -> 20.1Hurt
use BCE variant for short-caption retrieval (ablation)b-CLIP · MSCOCO T2IR@1: 34.0 -> 39.8Improved
infer with text-conditioned image embeddings instead of CLS (CE, K=6) (ablation)b-CLIP CE · Urban-1k T2IR@1: 87.9 -> 99.1Improved
infer with text-conditioned image embeddings at K=36 (BCE) (ablation)b-CLIP BCE · Urban-1k T2IR@1: 94.4 -> 88.9Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.

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