FlowCIR: Semantic Transport via Flow Matching
for Zero-Shot Composed Image Retrieval

Zhenqi He1, Ziqi Jiang1, Yuanpei Liu2, Yanghao Wang1, Teng Wang2, Long Chen1
1The Hong Kong University of Science and Technology 2The University of Hong Kong
Paper Project BibTeX
FlowCIR compares textual inversion with conditional semantic transport
Overview. Zero-shot CIR asks a model to retrieve a target image from a reference image and a relative instruction. FlowCIR moves beyond textual inversion by learning a reference-conditioned transport path in VLM embedding space.

Abstract

FlowCIR studies zero-shot composed image retrieval, where a target image is retrieved by applying a natural-language edit to a reference image without domain-specific CIR triplets. Instead of compressing the reference image into pseudo-text tokens, FlowCIR treats composition as conditional semantic transport: a lightweight flow-matching module transports the relative-instruction embedding toward a target-aligned query embedding under the reference-image condition. The image and text encoders stay frozen, making training efficient on pre-extracted VLM features. To address negation and removal queries, FlowCIR also introduces an inference-only Multi-Negative Steering strategy that pushes the instruction representation away from negated concepts before retrieval.

Highlights

Method

FlowCIR training and inference framework
Framework. During training, FlowCIR constructs a linear path between the relative-instruction embedding and the target-text embedding, then regresses the conditional velocity under the reference-image condition. During inference, a one-step transport produces a retrieval query for nearest-neighbor search.

Main Results

CIRR and CIRCO quantitative comparison
CIRR and CIRCO. FlowCIR achieves strong zero-shot CIR performance while using a much smaller training budget than many textual-inversion and generative baselines.

Qualitative Retrieval

FlowCIR qualitative retrieval examples
Qualitative examples. The same relative instruction can lead to different retrieved targets depending on the reference image. Multi-Negative Steering further improves retrieval behavior for removal-heavy queries.

BibTeX

@misc{he2026flowcir,
  title  = {FlowCIR: Semantic Transport via Flow Matching for Zero-Shot Composed Image Retrieval},
  author = {He, Zhenqi and Jiang, Ziqi and Liu, Yuanpei and Wang, Yanghao and Wang, Teng and Chen, Long},
  year   = {2026}
}