FlowComposer: Composable Flows for Compositional Zero-Shot Learning

TL;DR

FlowComposer introduces a flow-based framework for compositional zero-shot learning, explicitly modeling attribute-object compositions in the embedding space to improve generalization over prior PEFT-based methods.

Contribution

It is the first to systematically apply flow matching to CZSL, explicitly constructing compositions and addressing feature entanglement issues.

Findings

Achieves significant improvements on three CZSL benchmarks.

Effectively disentangles attribute and object features.

Enhances generalization in zero-shot composition recognition.

Abstract

Compositional zero-shot learning (CZSL) aims to recognize unseen attribute-object compositions by recombining primitives learned from seen pairs. Recent CZSL methods built on vision-language models (VLMs) typically adopt parameter-efficient fine-tuning (PEFT). They apply visual disentanglers for decomposition and manipulate token-level prompts or prefixes to encode compositions. However, such PEFT-based designs suffer from two fundamental limitations: (1) Implicit Composition Construction, where composition is realized only via token concatenation or branch-wise prompt tuning rather than an explicit operation in the embedding space; (2) Remained Feature Entanglement, where imperfect disentanglement leaves attribute, object, and composition features mutually contaminated. Together, these issues limit the generalization ability of current CZSL models. In this paper, we are the first to systematically study flow matching for CZSL and introduce FlowComposer, a model-agnostic framework that learns two primitive flows to transport visual features toward attribute and object text embeddings, and a learnable Composer that explicitly fuses their velocity fields into a composition flow. To exploit the inevitable residual entanglement, we further devise a leakage-guided augmentation scheme that reuses leaked features as auxiliary signals. We thoroughly evaluate FlowComposer on three public CZSL benchmarks by integrating it as a plug-and-play component into various baselines, consistently achieving significant improvements.