Research map / Vision-language and multimodal / CLIP and zero-shot
An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval
Jaeseok Byun, Seokhyeon Jeong, Wonjae Kim, Sanghyuk Chun, Taesup Moon
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 RTD text-encoder post-hoc training to projection-based CIR methods | Pic2Word, SEARLE, LinCIR, Context-I2W, FTI4CIR · CIRR/CIRCO/FashionIQ | average gain: average improvement of over 2.78 points | Improved |
| use generated text triplets (T_{r+c},T_t) instead of (T_r,T_r) pairs for contrastive loss (ablation) | RTD on LinCIR · CIR val (avg) | Avg: 38.5 -> 39.2 | Improved |
| add refined batch sampling with (T_r,T_r) hard negatives (ablation) | RTD on LinCIR · CIR val (avg) | Avg: 39.2 -> 39.6 | Improved |
| add refined concatenation scheme (ablation) | RTD on LinCIR · CIR val (avg) | Avg: 39.6 -> 40.7 | Improved |
| anchor target caption embedding with frozen text encoder (ablation) | RTD on LinCIR · CIR val (avg) | Avg: 36.3 -> 39.6 | Improved |
| naively fine-tune text encoder with original objective instead of RTD (ablation) | Pic2Word · CIR val (avg) | Avg: 32.2 -> 10.9 | Hurt |
| naively fine-tune text encoder with original objective (ablation) | LinCIR · CIR val (avg) | Avg: 36.9 -> 35.5 | Hurt |
| replace Compodiff LLM triplets with rule-based text triplets (ablation) | RTD on LinCIR · CIR val (avg) | Avg: +2.9 versus +3.3 over LinCIR 37.4 | No clear effect |
| use CASE text triplets (ablation) | RTD on LinCIR · CIR val (avg) | Avg: +0.3 | No clear effect |
| use CoVR text triplets (ablation) | RTD on LinCIR · CIR val (avg) | Avg: +4.5 | Improved |
| replace frozen text encoder with RTD-updated one in composed query | LinCIR ViT-L/14 · CIRCO val | cosine similarity to target image: 0.1 -> 0.29 | Improved |
| add RTD to projection-based methods | Pic2Word, SEARLE, LinCIR · GeneCIS | R@1: e.g. 12.45 -> 13.18 (LinCIR ViT-L); +0.10 for SEARLE ViT-L | Mixed |
Closest papers by meaning
- Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024 · similarity 0.94)
- ConText-CIR: Learning from Concepts in Text for Composed Image Retrieval (CVPR 2025 · similarity 0.94)
- Generative Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.93)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.92)
- Reason-before-Retrieve: One-Stage Reflective Chain-of-Thoughts for Training-Free Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.92)
- Vision-by-Language for Training-Free Compositional Image Retrieval (ICLR 2024 · similarity 0.92)
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.
