Research map / Vision-language and multimodal / CLIP and zero-shot
FlowCIR: Semantic Transport via Flow Matching for Zero-Shot Composed Image Retrieval
Zhenqi He, Ziqi Jiang, Yuanpei Liu, Yanghao Wang, Teng Wang, Long Chen
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 conditional flow matching on top of text-only query (ablation) | FlowCIR (ViT-L/14) · CIRCO test | mAP@5: 3.3 -> 13.5 | Improved |
| add conditional flow matching on top of text-only query (ablation) | FlowCIR (ViT-L/14) · CIRR test | R@5: 44.0 -> 53.8 | Improved |
| add Multi-Negative Steering on top of CFM (ablation) | FlowCIR (ViT-L/14) · CIRCO test | mAP@5: 13.5 -> 14.9 | Improved |
| add Multi-Negative Steering on top of CFM (ablation) | FlowCIR (ViT-L/14) · CIRR test | R@5: 53.8 -> 56.1 | Improved |
| add Multi-Negative Steering to text-only baseline (no CFM) (ablation) | Text-only query · CIRR and CIRCO test | R@5 / mAP@5: CIRR R@5 44.0 -> 45.7; CIRCO mAP@5 3.3 -> 2.5 | Mixed |
| replace flow matching with direct regression of target residual (ablation) | FlowCIR · CIRCO test | mAP@5: 14.9 -> 12.3 | Hurt |
| replace flow matching with direct regression of target residual (ablation) | FlowCIR · CIRR test | R@1: 26.2 -> 23.8 | Hurt |
| transport toward target image embedding instead of text embedding (ablation) | FlowCIR · CIRCO test | mAP@5: 14.9 -> 12.7 | Hurt |
| transport toward target image embedding instead of text embedding (ablation) | FlowCIR · CIRR test | R@1: 26.2 -> 23.8 | Hurt |
| raise retrieval loss weight lambda from 0.1 to 0.5 (ablation) | FlowCIR · CIRCO test | mAP@5: 10.6 -> 14.9 | Improved |
| raise retrieval loss weight lambda from 0.5 to 0.9 (ablation) | FlowCIR · CIRR/CIRCO test | R@1 / mAP@5: CIRR R@1 26.2 -> 25.8; CIRCO mAP@5 14.9 -> 14.8 | No clear effect |
| use hard-negative count K=B/8 instead of B/4 or B/16 (ablation) | FlowCIR · CIRCO test | mAP@5: 13.7 (B/4), 14.2 (B/16) -> 14.9 (B/8) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- Context-I2W — Context-I2W: Mapping Images to Context-Dependent Words for Accurate Zero-Shot Composed Image Retrieval (AAAI 2024) · better on 1 of 3 shared results
- iSEARLE — iSEARLE: Improving Textual Inversion for Zero-Shot Composed Image Retrieval (TPAMI 2025) · better on 2 of 2 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
- Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024 · similarity 0.89)
- Knowledge-Enhanced Dual-stream Zero-shot Composed Image Retrieval (CVPR 2024 · similarity 0.88)
- Reason-before-Retrieve: One-Stage Reflective Chain-of-Thoughts for Training-Free Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.88)
- Context-I2W: Mapping Images to Context-Dependent Words for Accurate Zero-Shot Composed Image Retrieval (AAAI 2024 · similarity 0.88)
- Image2Sentence based Asymmetrical Zero-shot Composed Image Retrieval (ICLR 2024 · similarity 0.87)
- Vision-by-Language for Training-Free Compositional Image Retrieval (ICLR 2024 · similarity 0.87)
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.
