Research map / Vision-language and multimodal / CLIP and zero-shot
Zero-Shot Composed Image Retrieval via Dual-Stream Instruction-Aware Distillation
Wenliang Zhong, Rob Barton, Weizhi An, Feng Jiang, Hehuan Ma, Yuzhi Guo, Abhishek Dan, Shioulin Sam, Karim Bouyarmane, Junzhou Huang
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 contrastive reasoning distillation loss with learnable prompt tokens to composed loss (ablation) | DistillCIR (ViT-L) · CIRR val R@1 | R@1: 30.38 -> 31.96 | Improved |
| use NLL text-generation loss for reasoning distillation instead of contrastive (ablation) | DistillCIR (ViT-L) · CIRR val R@1 | R@1: 30.38 -> 28.75 | Hurt |
| add learnable prompt tokens to the contrastive reasoning loss (ablation) | DistillCIR (ViT-L) · CIRR val R@1 | R@1: 31.40 -> 31.96 | Improved |
| add contrastive reasoning loss with prompt tokens (ablation) | DistillCIR (ViT-L) · CIRCO val | mAP@10: 19.75 -> 20.33 | Improved |
| add feature distillation from LLaVA-Phi3 MLLM teacher (ablation) | DistillCIR (ViT-L) · CIRR val R@1 | R@1: 30.38 -> 31.49 | Improved |
| use the smallest MLLM teacher (TinyLLaVA 1.4B) for feature distillation (ablation) | DistillCIR (ViT-L) · CIRR val R@1 | R@1: 30.38 -> 30.62 | No clear effect |
| use larger LLaVA-LLaMA3 (8B) teacher instead of LLaVA-Phi3 (4.2B) (ablation) | DistillCIR (ViT-L) · CIRR val R@1 | R@1: 31.49 -> 31.45 | No clear effect |
| DistillCIR vs CIReVL (LLM-based) on CIRCO test mAP@5 | DistillCIR (ViT-L) · CIRCO test | mAP@5: 18.57 -> 18.67 | No clear effect |
| DistillCIR vs CIReVL on FashionIQ Dress (ViT-B) | DistillCIR (ViT-B) · FashionIQ val Dress | R@10: 25.29 -> 25.14 | Hurt |
| DistillCIR vs Slerp on CIRR test R@1 (ViT-L) | DistillCIR (ViT-L) · CIRR test | R@1: 30.94 -> 32.34 | Improved |
Closest papers by meaning
- Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024 · similarity 0.93)
- Reason-before-Retrieve: One-Stage Reflective Chain-of-Thoughts for Training-Free Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.93)
- Vision-by-Language for Training-Free Compositional Image Retrieval (ICLR 2024 · similarity 0.92)
- Image2Sentence based Asymmetrical Zero-shot Composed Image Retrieval (ICLR 2024 · similarity 0.92)
- Knowledge-Enhanced Dual-stream Zero-shot Composed Image Retrieval (CVPR 2024 · similarity 0.91)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.91)
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.
