FlowComposer: Composable Flows for Compositional Zero-Shot Learning

Zhenqi He; Lin Li; Long Chen

cs.CVMar 17, 2026

FlowComposer: Composable Flows for Compositional Zero-Shot Learning

Authors: Zhenqi He, Lin Li, Long Chen

Organizations: The Hong Kong University of Science and Technology

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.

Figures & tables

Panel of Figure 1 — Figure 1: (a) Humans recognize new concepts by recombining familiar primitives. (b) Prior CZSL methods compose only at the token level, which may not yield valid unseen compositions in the embedding space. (c) We perform explicit composition in the embedding space via learned attribute and object flows.

Figure 2: Training dynamics and performance comparison with baseline - Troika [ 15 ] . Our method yields a more balanced seen/unseen accuracy trajectory and consistently improves HM and AUC over the baseline on all three datasets.

Figure 3: Overall framework for FlowComposer . (a) Primitive flows: Two flow models learn time–conditioned velocities to transport primitive visual embeddings to their corresponding text embeddings. (b) Composer: the network to predict the combination coefficients supervised by least-squares targets. (c) Leakage-Guided Augmentation: To exploit residual cross-branch cues, each primitive flow is also trained to transport leaked features from the counterpart (or composition) branch to its own text target.

Closed-world Results
	MIT-States [ 17 ]				UT-Zappos [ 52 ]				C-GQA [ 39 ]
Method	Seen	Unseen	HM	AUC	Seen	Unseen	HM	AUC	Seen	Unseen	HM	AUC
CLIP [ 45 ] [ICML21]	30.2	46.0	26.1	11.0	15.8	49.1	15.6	5.0	7.5	25.0	8.6	1.4
CoOP [ 55 ] [IJCV22]	34.4	47.6	29.8	13.5	52.1	49.3	34.6	18.8	20.5	26.8	17.1	4.4
DFSP(i2t) [ 36 ] [CVPR23]	47.4	52.4	37.2	20.7	64.2	66.4	45.1	32.1	35.6	29.3	24.3	8.7
DFSP(BiF) [ 36 ] [CVPR23]	47.1	52.8	37.7	20.8	63.3	69.2	47.1	33.5	36.5	32.0	26.2	9.9

Table 1: Quantitative comparison (§ 5.2 ) on three benchmarks within closed-world and open-world setting. * denotes results from our implementation.

\ddagger

represents the methods leveraging the LLM’s knowledge.

	Flows	Composer	LG-Aug	MIT-States [ 17 ]				UT-Zappos [ 52 ]
	(§ 4.1 )	(§ 4.2 )	(§ 4.3 )	Seen	Unseen	HM	AUC	Seen	Unseen	HM	AUC
Baseline	✗	✗	✗	49.3	52.5	39.2	22.1	66.3	73.4	55.4	41.8
(1)	✓	✗	✗	51.3	52.1	39.4	22.6	68.8	74.8	55.1	43.0
(2)	✓	✗	✓	51.3	53.0	40.3	23.3	69.2	74.9	56.3	44.0
(3)	✓	✓	✗	51.0	53.0	40.1	23.3	69.7	74.9	58.5	45.8
Ours	✓	✓	✓	51.5	53.2	40.2	23.5	71.1	74.9	58.6	46.8

Table 2: Ablations. The results regarding the different components in our FlowComposer on MIT-States [ 17 ] and UT-Zappos [ 52 ] .

	MIT-States				UT-Zappos
	S	U	H	A	S	U	H	A
N/A	49.3	52.5	39.2	22.1	66.3	73.4	55.4	41.8
c	50.9	48.0	36.9	20.0	70.0	60.1	44.6	37.2
c+a+o	51.0	48.7	37.4	20.5	70.1	60.6	44.9	37.6
FlowComposer	51.5	53.2	40.2	23.5	71.1	74.9	58.6	46.8

Table 3: Analysis on flows (§ 4.1 ) on the MIT-States [ 17 ] and UT-Zappos [ 52 ] under closed-world setting.

Figure 4: Visual comparisons between CSP [ 41 ] , Troika [ 15 ] and our FlowComposer on three datasets. Red represents the wrong prediction, and Green represents the right prediction.

	MIT-States		UT-Zappos		C-GQA
	HM	AUC	HM	AUC	HM	AUC
Predictor	39.9	23.3	58.6	45.2	33.9	15.7
Composer	40.2	23.5	58.6	46.8	34.0	15.9

Table 4: Comparison between Predictor and Composer in MIT-States [ 17 ] , UT-Zappos [ 52 ] and C-GQA [ 39 ] on HM and AUC under the closed-world setting.

	MIT-States		UT-Zappos		C-GQA
	HM	AUC	HM	AUC	HM	AUC
Regressor	39.6	22.5	55.2	42.8	30.0	12.9
FM	40.2	23.5	58.6	46.8	34.0	15.9

Table 5: Comparison between regressor and our FM on HM and AUC under the closed-world setting.

			Training		Validation			Test
Dataset	$\mathcal{S}$	$\mathcal{O}$	$\mathcal{C}^{se}$	$\mathcal{X}$	$\mathcal{C}^{se}$	$\mathcal{C}^{us}$	$\mathcal{X}$	$\mathcal{C}^{se}$	$\mathcal{C}^{us}$	$\mathcal{X}$
MIT-States [ 17 ]	115	245	1262	30k	300	300	10k	400	400	13k
UT-Zappos [ 52 ]	16	12	83	23k	15	15	3k	18	18	3k
C-GQA [ 39 ]	413	674	5592	27k	1252	1040	7k	888	923	5k

Table 6: Statistics of three datasets in our experiments. The number of elements in each set is reported.

Network	# Params (M) $\downarrow$	GFLOPs $\downarrow$
FM	74.06	0.07
Composer	25.97	0.03

Table 7: Model size and computational cost of our networks.

	MIT-States		UT-Zappos		C-GQA
Acc ( $\%$ )	Attr	Obj	Attr	Obj	Attr	Obj
Random	1.13	0.47	5.74	8.14	0.25	0.05
Attr Branch	26.3	3.87	22.3	14.9	3.64	2.36
Obj Branch	6.3	45.2	7.69	63.2	2.12	17.3
Comp Branch	21.3	51.3	21.6	58.9	9.72	39.8

Table 8: Comparison between random prediction and predictions from Attribute , Object and Composition branches on class-balanced accuracy ( Acc ) for attributes and objects.

	MIT-States		UT-Zappos		C-GQA
Disentangler	HM	AUC	HM	AUC	HM	AUC
w	39.2	22.1	55.4	41.8	29.7	12.5
w/o	36.8	20.4	51.4	37.5	28.0	11.2

Table 9: Comparison between Troika and Troika without disentangler in MIT-States [ 17 ] , UT-Zappos [ 52 ] and C-GQA [ 39 ] on HM and AUC under the closed-world setting.

	Flows	Composer	LG-Aug	MIT-States [ 17 ]				UT-Zappos [ 52 ]				C-GQA [ 39 ]
	(§ 4.1)	(§ 4.2)	(§ 4.3)	Seen	Unseen	HM	AUC	Seen	Unseen	HM	AUC	Seen	Unseen	HM	AUC
Baseline	✗	✗	✗	50.3	17.5	19.0	6.8	65.8	61.0	46.6	32.7	40.8	8.6	11.7	2.9
(1)	✓	✗	✗	50.1	18.0	19.6	7.0	68.5	60.8	49.3	33.8	41.2	9.0	12.0	3.2
(2)	✓	✗	✓	49.7	18.9	20.2	7.3	69.1	61.5	50.1	34.7	42.2	10.0	12.3	3.3
(3)	✓	✓	✗	49.6	18.8	20.0	7.2	69.8	61.1	50.7	34.9	41.9	10.3	12.2	3.2
Ours	✓	✓	✓	50.4	19.0	20.3	7.5	70.1	61.2	51.0	35.5	43.5	10.2	12.6	3.5

Table 10: Ablations. The results regarding the different components in our FlowComposer on MIT-States [ 17 ] , UT-Zappos [ 52 ] and C-GQA [ 39 ] under the open-world setting.

Method	MIT-States	UT-Zappos	C-GQA
CSP	6.6	10.3	20.2
+ Ours	9.3	12.5	24.6
Troika	17.6	11.0	69.8
+ Ours	19.2	12.7	74.4

Table 11: Inference time (ms/image) comparison between two baselines and ours.

	Validation				Test
$h$	S	U	H	A	S	U	H	A
0.1	44.6	40.6	32.9	15.5	44.8	40.7	34.0	15.9
0.5	44.6	39.9	32.6	15.2	43.8	40.0	32.9	15.1
1.0	42.0	36.7	29.8	13.0	40.3	36.9	29.7	12.7

Table 12: Results regarding stepsize

h

(§ 4.2) on the validation and test set under the closed-world setting for C-GQA [ 39 ] .

	MIT-States [ 17 ]				UT-Zappos [ 52 ]				C-GQA [ 39 ]
Method	Seen	Unseen	HM	AUC	Seen	Unseen	HM	AUC	Seen	Unseen	HM	AUC
CSP ∗ [ 41 ] [ICLR23]	46.6 $\pm 0.1$	49.9 $\pm 0.1$	36.3 $\pm 0.1$	19.4 $\pm 0.1$	64.2 $\pm 0.7$	66.2 $\pm 1.2$	46.6 $\pm 1.2$	33.0 $\pm 1.3$	28.8 $\pm 0.1$	26.8 $\pm 0.1$	20.5 $\pm 0.1$	6.2 $\pm 0.0$
+FlowComposer	48.4 $\pm 0.2$	50.4 $\pm 0.1$	37.8 $\pm 0.2$	20.7 $\pm 0.1$	66.9 $\pm 0.4$	68.0 $\pm 0.7$	51.4 $\pm 0.7$	38.2 $\pm 0.8$	29.0 $\pm 0.1$	31.0 $\pm 0.2$	23.1 $\pm 0.2$	7.8 $\pm 0.2$
Troika ∗ [ 15 ] [CVPR24]	49.0 $\pm 0.4$	53.0 $\pm 0.2$	39.3 $\pm 0.2$	22.1 $\pm 0.1$	66.8 $\pm 1.1$	73.8 $\pm 0.6$	54.6 $\pm 0.5$	41.7 $\pm 0.7$	41.0 $\pm 0.2$	35.7 $\pm 0.3$	29.7 $\pm 0.2$	12.4 $\pm 0.1$
+FlowComposer	51.7 $\pm 0.3$	53.1 $\pm 0.1$	40.2 $\pm 0.1$	23.4 $\pm 0.1$	71.5 $\pm 0.6$	75.0 $\pm 0.3$	58.8 $\pm 0.4$	46.7 $\pm 0.3$	44.8 $\pm 0.1$	40.6 $\pm 0.1$	34.1 $\pm 0.1$	15.9 $\pm 0.1$

Table 13: Performance on MIT-States [ 17 ] , UT-Zappos [ 52 ] , and C-GQA [ 39 ] , averaged over 5 random seeds with corresponding standard deviations. * marks results taken from the respective original publications.

Figure 5: The purple bar denotes the value of $\hat{a}$ and the green bar denotes the value of $\hat{b}$ , which are the coefficients for attribute and object velocities respectively.

Zero-Shot Learning

Zero-shot learning lets models recognize classes or solve tasks absent from task-specific training examples by using semantic descriptions or pretrained knowledge.

See all 184 papers

DIFFCZSL: Compositional Zero-Shot Learning Regularized by Diffusion Representations
Aug 20, 2026Hangyu Tian, Zhenqi He, Yanghao Wang +1Zero-Shot Learning Compositional Representation Learning
The Hong Kong University of Science and Technology
FlowCIR: Semantic Transport via Flow Matching for Zero-Shot Composed Image Retrieval
Jul 2, 2026Zhenqi He, Ziqi Jiang, Yuanpei Liu +3Vision-Language Models Composed Image Retrieval
The Hong Kong University of Science and Technology, Hong Kong SAR · The University of Hong Kong, Hong Kong SAR
Compositionality with Primitives: A Unified View from Euclidean to Hyperbolic Vision-Language Models
Oct 21, 2026Puru Vaish, Meelad Dashti, Christoph Brune +3Compositional Reasoning in VLMs Zero-Shot Learning
University of Twente, The Netherlands · Politecnico di Torino, Italy · University of Amsterdam, The Netherlands

Aug 20, 2026J/K move · Enter open · S save

DIFFCZSL: Compositional Zero-Shot Learning Regularized by Diffusion Representations

Hangyu Tian, Zhenqi He, Yanghao Wang, Long Chen

The Hong Kong University of Science and Technology

Zero-Shot Learning Compositional Representation Learning

Open paper AskPDF arXiv

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen attribute-object compositions by leveraging knowledge of primitive concepts learned from seen compositions. Although recent works achieve impressive performance in CZSL by leveraging large vision-language models, they primarily rely on discriminative representations that may not explicitly preserve the structured relationships between primitive concepts and their compositions. Motivated by the recent success of diffusion-based classifiers and their competitive performance relative to discriminative models, we investigate whether intermediate diffusion representations can provide complementary cues for CZSL. To this end, we propose DIFFCZSL, a diffusion-augmented framework that injects generative priors from pre-trained diffusion models into CLIP-based CZSL pipelines. We extract intermediate diffusion representations and project them into the CLIP embedding space to provide auxiliary supervision on both image and text modalities. Through contrastive alignment between CLIP embeddings and diffusion features during training, our method encourages the embedding geometry toward richer composition-aware semantics, while introducing no additional cost at inference time. Extensive experiments on three public CZSL benchmarks demonstrate consistent improvements over strong CLIP-based baselines under both closed-world and open-world settings. Our results highlight the complementary strengths of generative diffusion representations and discriminative vision-language models for compositional generalization.

Abstract

Zero-Shot Learning

Explore similar work

DIFFCZSL: Compositional Zero-Shot Learning Regularized by Diffusion Representations

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

Compositionality with Primitives: A Unified View from Euclidean to Hyperbolic Vision-Language Models