DiffReID: Discriminative Diffusion Model for Object Re-Identification

Yingquan Wang; Pingping Zhang; Dong Wang; Huchuan Lu

cs.CVSep 29, 2026

DiffReID: Discriminative Diffusion Model for Object Re-Identification

Authors: Yingquan Wang, Pingping Zhang, Dong Wang, Huchuan Lu

Organizations: School of Information and Communication Engineering, Dalian University of Technology · School of Future Technology, Dalian University of Technology

Abstract

As a fundamental image processing task, object Re-Identification (ReID) aims to retrieve objects across non-overlapping cameras. Recently, with the development of deep learning, significant advancements have been made in object ReID. However, most existing methods suffer from generalization due to the limited size and diversity of ReID datasets. Meanwhile, current models tend to focus on extracting semantic patterns rather than learning identity-aware feature distributions. To address these issues, we propose a novel feature learning framework named \textbf{DiffReID} for object ReID. It leverages a discriminative diffusion model to gradually learn identity-aware distributions and generate identity-invariant features. More specifically, with the Contrastive Language-Image Pre-training (CLIP) model, we first obtain identity-aware text features by prompt tuning. Then, we propose a Vision-guided Noise Generator (VNG) to initialize probabilistic noises and gradually corrupt identity-aware text features. Afterwards, we take visual features as conditions and propose a Light Weight Denoiser (LWD) to denoise the corrupted text features step-by-step for identity-aware distribution learning. To obtain discriminative features, we further generate identity-invariant guided features from randomly sampling visual-guided noises. Finally, we propose a Mutual Enhancement Constraint (MEC) to facilitate mutual learning between visual features and guided features to enhance the representation robustness and discrimination. Extensive experiments on five object ReID benchmarks demonstrate that our method shows better results than most state-of-the-art methods. The source code is available at https://github.com/AWangYQ/DiffReID.

Figures & tables

Fig. 1 : Different diffusion model frameworks for object ReID. (a) The refinement diffusion model framework unifies feature extraction and denoising within a backbone pipeline. (b) The restoration diffusion model framework utilizes intermediate features from diffusion models for object representation. (c) Our discriminative diffusion model framework generates identity-invariant guided features and facilitates mutual learning between guided features and visual features.

Fig. 2 : Illustration of our proposed framework. With the CLIP model, the framework first generates identity-aware text features by prompt tuning. Then, the Vision-guided Noise Generator (VNG) is used to initialize probabilistic noises and gradually corrupt identity-aware text features. Afterwards, the Light Weight Denoiser (LWD) is used to denoise the corrupted text features for identity-aware distribution learning. To train the framework, the Mutual Enhancement Constraint (MEC) is performed to facilitate mutual learning between visual features and guided features.

Method	Market1501		MSMT17		DukeMTMC		Method	VeRi-776
Method	mAP	Rank-1	mAP	Rank-1	mAP	Rank-1	Method	mAP	Rank-1
Nformer [ 42 ]	91.1	94.7	59.8	77.3	83.5	89.4	CFVMNet [ 43 ]	77.1	95.3
CMT [ 27 ]	87.7	95.4	62.6	83.1	80.0	90.1	DCAL [ 44 ]	80.2	96.9
TransReID [ 1 ]	88.9	95.2	67.4	85.3	82.0	90.7	GLTrans [ 4 ]	82.9	97.5
GLTrans [ 4 ]	90.0	95.6	69.0	85.8	82.4	90.7	HPGN [ 45 ]	80.2	96.7
HAT [ 24 ]	89.8	95.8	61.2	82.3	81.4	90.4	PGAN [ 46 ]	79.3	96.5

TABLE I : Comparison of state-of-the-art methods on three person ReID datasets and one vehicle ReID dataset. The best and second results are marked in bold and underlined , respectively.

Method	Training	Market1501		MSMT17		CUHK03-NP		DukeMTMC
Method	Training	mAP	R1	mAP	R1	mAP	R1	mAP	R1
QAConv [ 70 ]	Market1501	–	–	7.0	22.6	8.6	9.9	33.6	54.4
TransMatcher [ 71 ]		–	–	18.4	47.3	21.4	22.2	–	–
QAConv+Gs [ 72 ]		–	–	17.2	45.9	18.1	19.1	–	–
MDA [ 67 ]		–	–	11.8	33.5	–	–	34.4	56.7
PAT [ 73 ]		–	–	18.2	42.8	26.0	25.4	48.9	67.9

TABLE II: Comparison of state-of-the-art methods on domain generalization settings. The best and second results are marked in bold and underlined , respectively. The superscript * indicates that the model resizes the input images to 384

\times

128.

	Baseline	LWD	VNG	MEC	MSMT17		DukeMTMC
	Baseline	LWD	VNG	MEC	mAP	R1	mAP	R1
(a)	✓	$\times$	$\times$	$\times$	75.4	89.3	82.8	90.7
(b)	✓	✓	$\times$	$\times$	76.6	90.4	83.9	91.1
(c)	✓	✓	✓	$\times$	77.7	90.8	84.5	91.7
(d)	✓	✓	$\times$	✓	77.2	90.9	84.6	92.0
(e)	✓	✓	$\checkmark$	$\checkmark$	79.0	91.2	85.6	92.4

TABLE III : Ablation results with different components.

No.	MSMT17		DukeMTMC
No.	mAP	Rank-1	mAP	Rank-1
Baseline	75.4	89.3	82.8	90.7
2	77.0	90.2	84.4	92.1
4	77.5	90.5	84.6	92.3
6	78.0	90.9	85.0	92.0
8	78.6	91.0	85.3	92.2

TABLE IV : Comparison with different diffusion steps.

MEC	Feature	Performance		Cosine Similarity		Generation Consistency $\uparrow$
MEC	Feature	mAP	Rank-1	Intra-class $\uparrow$	Inter-class $\downarrow$	Generation Consistency $\uparrow$
w/o	Visual	77.0	89.5	0.72	0.35	–
w/o	Guided	77.7	90.8	0.76	0.29	0.75
w/	Visual	77.8	90.3	0.78	0.29	–
w/	Guided	79.0	91.2	0.83	0.24	0.89

TABLE V : Effect of MEC with visual and guided features.

No.	MSMT17		DukeMTMC
No.	mAP	Rank-1	mAP	Rank-1
Baseline	75.4	89.3	82.8	90.7
1	76.7	90.5	84.1	91.7
2	78.1	90.9	84.8	92.0
3	79.0	91.2	85.6	92.4
4	79.0	91.3	85.4	92.2

TABLE VI : Comparison with different guided features.

No.	MSMT17		DukeMTMC
No.	mAP	Rank-1	mAP	Rank-1
2	76.7	89.9	83.9	91.4
4	77.1	90.1	84.1	91.6
8	77.5	90.6	84.3	92.1
16	79.0	91.2	85.6	92.4
32	78.7	91.0	85.4	92.4

TABLE VII : Effect of different number of temporal queries.

Setting	MSMT17		DukeMTMC
Setting	mAP	Rank-1	mAP	Rank-1
$f_{12}$	77.8	90.8	84.4	91.0
$f_{12}+f_{11}$	78.0	91.2	84.5	91.2
$f_{12}+f_{11}+f_{10}$	78.6	91.4	84.6	91.5
$f_{12}+f_{11}+f_{10}+f_{9}$	78.8	91.3	84.9	91.7
$f_{12}+f_{11}+f_{10}+f_{9}+f_{8}$	78.8	91.0	85.1	92.0

TABLE VIII : Effect of intermediate features for the generation.

Model	Seed	MSMT17		DukeMTMC
Model	Seed	mAP	Rank-1	mAP	Rank-1
Baseline	1234	75.4	89.3	82.8	90.7
DiffReID	1	78.8	91.3	85.8	92.3
DiffReID	12	78.4	91.4	85.5	92.0
DiffReID	123	78.7	90.7	85.4	92.2
DiffReID	1234	79.0	91.2	85.6	92.4

TABLE IX : Effect of different random seeds.

Setting	MSMT17		DukeMTMC
Setting	mAP	Rank-1	mAP	Rank-1
Baseline	75.4	89.3	82.8	90.7
On visual features	77.8	90.5	84.5	91.8
On identity-aware text features	79.0	91.2	85.6	92.4

TABLE X : Effect of different diffusion features.

Model	MSMT17		DukeMTMC
Model	mAP	Rank-1	mAP	Rank-1
Baseline	75.4	89.3	82.8	90.7
DiffReID w/o KL	77.6	90.8	83.7	91.1
DiffReID	79.0	91.2	85.6	92.4

TABLE XI : Effect of the KL regularization in VNG.

Method	Trainable Params (M)	Memory (GB)	GFLOPs	Time (s/batch)	mAP (MSMT17)
SD-ReID [ 78 ]	103.36	11.05	677.67	1.66	–
FusionReID [ 28 ]	153.80	1.62	28.10	0.14	69.5
TF-CLIP [ 3 ]	104.26	1.53	24.24	0.26	73.9
DiffReID	104.81	1.32	43.40	0.22	79.0

TABLE XII : Cost comparison with different methods.

Fig. 3 : Retrieval results on MSMT17. Black, green and red boxes indicate query images, correct matches and incorrect matches, respectively.

Fig. 4 : t-SNE visualization of guided features at successive denoising timesteps.

Fig. 5 : Feature distributions visualized by t-SNE. Different colors represent different identities.

Fig. 6 : Visualization of the cosine similarity distribution.

Fig. 7 : Visualization of attention maps for different temporal queries. Darker red indicates higher attention weights.

Explore similar work

Apr 21, 2026cs.CV

Thinking Before Matching: A Reinforcement Reasoning Paradigm Towards General Person Re-Identification

Learning identity-discriminative representations with multi-scene generality has become a critical objective in person re-identification (ReID). However, mainstream perception-driven paradigms tend to identify fitting from massive annotated data rather than identity-causal cues understanding, which presents a fragile representation against multiple disruptions. In this work, ReID-R is proposed as a novel reasoning-driven paradigm that achieves explicit identity understanding and reasoning by incorporating chain-of-thought into the ReID pipeline. Specifically, ReID-R consists of a two-stage contribution: (i) Discriminative reasoning warm-up, where a model is trained in a CoT label-free manner to acquire identity-aware feature understanding; and (ii) Efficient reinforcement learning, which proposes a non-trivial sampling to construct scene-generalizable data. On this basis, ReID-R leverages high-quality reward signals to guide the model toward focusing on ID-related cues, achieving accurate reasoning and correct responses. Extensive experiments on multiple ReID benchmarks demonstrate that ReID-R achieves competitive identity discrimination as superior methods using only 14.3K non-trivial data (20.9% of the existing data scale). Furthermore, benefit from inherent reasoning, ReID-R can provide high-quality interpretation for results.

Quan Zhang, Jingze Wu, Jialong Wang +3

Sun Yat-sen University · Alibaba Cloud Computing

Re-Identification Scene Reasoning Fine-Grained Perception Discriminative

Ask PDF

Jul 31, 2026cs.CV

Multi-Modal Object Re-Identification with Dual Semantic Guidance and Global-Local Mutual Modulation

Multi-modal object Re-Identification (ReID) aims to retrieve target instances by leveraging complementary information across modalities. However, existing methods suffer from two challenges. First, they often fail to exploit well-aligned and reliable semantic priors, making them vulnerable to background clutter and cross-modal misalignment. On the other hand, they typically rely on holistic feature modeling, overlooking the synergy between global and local representations. To overcome these limitations, we propose a robust multi-modal ReID framework with dual semantic guidance and global-local mutual modulation, which mainly consists of three key components, namely the Text-Semantic Injector (TSI), the Masked Global-Local Modulator (MGLM), and the Hierarchical MoE Fusion (HMF). The TSI enhances semantic awareness by integrating clean and coherent textual features into visual tokens. The MGLM enables part-aware cross-modal interaction through joint guidance from soft masks and global context, improving fine-grained feature alignment. Finally, the HMF adaptively aggregates multi-spectral features under local semantic supervision, yielding discriminative and robust representations. Extensive experiments on three multi-modal ReID benchmarks demonstrate the effectiveness of the proposed method. The code will be made publicly available at https://github.com/zw-absin/DSGM upon acceptance.

Weixiang Zhou, Xingguo Xu, Yuhao Wang +4

School of Mathematical Sciences and the Key Laboratory for Computational Mathematics and Data Intelligence of Liaoning Province, Dalian University of Technology, Dalian, 116024, China · School of Future Technology, School of Artificial Intelligence, Dalian University of Technology, Dalian, 116024, China · Department of Radiology and Biomedical Imaging, University of California, San Francisco, 94107, USA +1

Re-Identification

Ask PDF

Sep 21, 2026cs.CV

Incentive Noise and Structural Prior Infusion for Multi-modal Object Re-Identification

Multi-modal object Re-Identification (ReID) benefits from complementary information across heterogeneous imaging modalities. To further enrich semantic representation, text descriptions have recently been incorporated as an additional modality. However, recent vision-language approaches often treat text descriptions as clean, deterministic signals and overlook their inherent noise, including modality-mismatched phrases and semantically ambiguous expressions. Moreover, prevailing methods lack explicit mechanisms to reconcile fine-grained structural discrepancies between modalities, even after high-level semantic alignment. To address these challenges, we propose a novel framework centered on Positive-Incentive Noise (π-noise) and structured prompt modulation. First, the Semantic Cross-Modal Modulator harnesses task-aware π-noise, sampled from a distribution conditioned on both visual and text inputs, to perturb global tokens and enable semantics-guided cross-modal compensation. Second, the Structure-Aware Prompt Adapter injects learnable geometric priors via prompts to enhance spatial consistency. Third, the Context-Aware Sparse Fusion module distills structural context to guide adaptive fusion while shielding identity features from noisy local details. Experiments on three multi-modal ReID benchmarks demonstrate the effectiveness and robustness of our approach. The code is available at https://github.com/zw-absin/INSPI.

Weixiang Zhou, Yuhao Wang, Xingguo Xu +4

Dalian University of Technology, China · New York University, USA · Nanjing University of Science and Technology, China +1

Re-Identification Cross-Modal Structure-Aware Incentives

Ask PDF