Research map / Vision-language and multimodal / CLIP and zero-shot
Spherical Linear Interpolation and Text-Anchoring for Zero-shot Composed Image Retrieval
Young Kyun Jang, Dat B Huynh, Ashish Shah, Wen-Kai Chen, Ser-Nam Lim
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add TAT (text-anchored LoRA tuning of image encoder) before Slerp | Slerp ZS-CIR · CIRCO test, CLIP-ViT-L/14 | mAP@5: 8.76 -> 18.46 | Improved |
| add TAT before Slerp | Slerp ZS-CIR · CIRR test, CLIP-ViT-L/14 | Recall@1: 24.43 -> 30.94 | Improved |
| fix text encoder and tune image encoder (text-anchoring) rather than tune both encoders (ablation) | TAT · CIRCO test, CLIP-ViT-L/14 | mAP@5: 8.26 -> 18.46 | Improved |
| fix text encoder and tune image encoder rather than image-anchoring (ablation) | TAT · CIRCO test, CLIP-ViT-L/14 | mAP@5: 7.54 -> 18.46 | Improved |
| scale TAT training data from Laion-2M to Laion-4M or 8M (ablation) | TAT · CIRCO test, CLIP-ViT-L/14 | mAP@5: 18.46 -> 18.45 (4M), 18.23 (8M) | No clear effect |
| train TAT on only 0.58M LLaVA-Align pairs (ablation) | TAT · CIRCO test, CLIP-ViT-L/14 | mAP@5: 17.05 vs 18.46 with Laion-2M | No clear effect |
| train Pic2Word on Laion-2M instead of CC3M (ablation) | Pic2Word · CIRCO test, CLIP-ViT-L/14 | mAP@5: 8.93 vs Pic2Word 8.72 | No clear effect |
| apply LoRA to Pic2Word (ablation) | Pic2Word · CIRCO test, CLIP-ViT-L/14 | mAP@5: 8.89 vs Pic2Word 8.72 | No clear effect |
| raise Slerp text weight alpha up to 0.8 then to 1.0 (no image) (ablation) | Slerp · CIRCO test, CLIP-ViT-L/14 | mAP and Recall | Mixed |
| replace original CLIP with TAT-trained CLIP in supervised Combiner training | Combiner · CIRR test, CLIP-ViT-L/14 | Recall@1: 30.31 -> 36.58 | Improved |
| replace original CLIP with TAT-trained CLIP in supervised Combiner training | Combiner · FashionIQ val, CLIP-ViT-L/14 | Average R@10: 35.64 -> 36.69 | Improved |
| add TAT on top of Slerp | Slerp ZS-CIR · FashionIQ val, CLIP-ViT-B/32 | Average R@10: 23.75 -> 22.96 | Mixed |
Closest papers by meaning
- Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024 · similarity 0.93)
- ConText-CIR: Learning from Concepts in Text for Composed Image Retrieval (CVPR 2025 · similarity 0.92)
- Generative Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.91)
- Reason-before-Retrieve: One-Stage Reflective Chain-of-Thoughts for Training-Free Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.91)
- Vision-by-Language for Training-Free Compositional Image Retrieval (ICLR 2024 · similarity 0.90)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.90)
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.
