LReID-Hybrid: Lifelong Person ReID

• The LReID-Hybrid task is a continual re-ID setting where models update sequentially on same-cloth and cloth-changing domains without relying on data replay.
• It employs cross-modal strategies such as text-space alignment, prompt learning, and prototype accumulation to mitigate catastrophic forgetting and address clothing-state mismatch.
• Benchmarks using mAP and Rank-1 show robust performance on same-cloth domains while highlighting persistent challenges in cloth-changing and unseen domains.

LReID-Hybrid denotes lifelong person re-identification with hybrid clothing states, a continual ReID setting in which a model learns from a sequence of domains that may be either same-cloth (SC) or cloth-changing (CC), while preserving performance on previously seen domains and generalizing to unseen ones. In the formulation introduced for this task, the learner receives only the current domain at each step, cannot rely on data replay, and must cope simultaneously with catastrophic forgetting and clothing-state variation. The central methodological theme across the literature is that SC and CC domains induce mismatched supervision signals and representation biases, so effective systems increasingly use cross-modal structure—especially text-space alignment, prompt learning, prototype accumulation, or distribution rehearsal—to bridge old and new domains without storing raw historical images (Wang et al., 2024).

1. Formal problem setting

A standard formulation considers a sequence of ReID domains

$$\mathcal D = \bigl\{\mathcal D^{(1)}, \mathcal D^{(2)}, \ldots, \mathcal D^{(T)}\bigr\},$$

where each domain

$$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$

is either a Same-Cloth domain, in which every identity wears a single clothing style across camera views, or a Cloth-Changing domain, in which identities may appear in different outfits. At step $t$, the model has access only to $\mathcal D_{\rm train}^{(t)}$, updates the image encoder from $\mathcal G_{t-1}$ to $\mathcal G_t$, and is evaluated after the final step on all seen domains $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$ as well as unseen domains $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$ (Wang et al., 2024).

A later CLIP-based formulation expresses the same objective with explicit clothing-state labels. At stage $t$, the model observes

$$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$

where $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$0, and updates $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$1 to $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$2 while maintaining retrieval performance on all previously seen domains. Evaluation is reported as SC-Average, CC-Average, and Total-Average in terms of mean Average Precision (mAP) and Rank-1 (Long et al., 14 Sep 2025).

This task should be distinguished from conventional CCReID and from standard LReID. CCReID isolates clothing change as a standalone retrieval problem; LReID emphasizes continual adaptation across domains; LReID-Hybrid makes both factors active at once. A plausible implication is that anti-forgetting mechanisms that work for same-cloth streams may fail when the incoming domain changes the semantic status of apparel cues.

2. Why hybrid clothing states are difficult

The original motivation for LReID-Hybrid is that existing LReID studies accumulate knowledge with the assumption that people would not change their clothes, whereas practical surveillance systems contain both cloth-consistent and cloth-changing domains (Wang et al., 2024). Two failure modes are emphasized: knowledge granularity mismatch and knowledge presentation mismatch. In SC data, clothing often behaves as a stable identity cue; in CC data, the same cue becomes unreliable or even misleading. As a result, the model must preserve identity information while dynamically reweighting clothing-sensitive and clothing-invariant evidence.

The multimodal literature around person retrieval makes the same tension visible from another angle. Instruct-ReID treats “traditional ReID,” “clothes-changing,” “visible–infrared,” “text-to-image,” and “language-instructed” ReID as special cases of a unified instruction-driven retrieval problem, where a query image or template is fused with an instruction and matched against a gallery. The accompanying OmniReID benchmark aggregates twelve public datasets across six formulations and yields approximately $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$3M training images of $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$4k identities, with evaluation on held-out test splits without fine-tuning (He et al., 2023). This suggests that clothing-state variation is not an isolated nuisance factor but one axis of a broader instruction-conditioned retrieval space.

A common misconception is that “hybrid” here refers only to multimodal input. In the LReID-Hybrid literature, the primary meaning is the coexistence of SC and CC domains in a lifelong stream. Multimodality becomes important because text space, prompts, or image–text alignment can provide a more stable intermediate representation than raw appearance alone.

3. Teata and the “image-text-image” closed loop

The first dedicated framework for LReID-Hybrid is Teata, which uses the consistency and generalization of the text space to align, transfer, and accumulate knowledge in an “image-text-image” closed loop (Wang et al., 2024). Its optimization objective is written as

$$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$5

subject to “no data replay,” with a frozen CLIP text encoder $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$6 and prompt parameters $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$7 carried forward across tasks.

Stage 1 is Structured Semantic Prompt learning. For each identity $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$8 in domain $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$9, Teata constructs

$t$0

where the $t$1 are shared tokens and the $t$2 are specific tokens. The previous image encoder $t$3 and the CLIP text encoder are frozen, and symmetric image-to-text and text-to-image contrastive losses align image features $t$4 and prompt-derived text features $t$5. During this stage, only the prompt tokens are updated.

Stage 2 is Knowledge Adaptation and Projection. The text embeddings become identity-level semantic prototypes, are L2-normalized, and are then adapted by a slow-paced learner while the image encoder is updated. Training uses projection-based ID loss, adapted-text distillation loss, image-space ID loss, and triplet loss, combined as

$t$6

with $t$7.

On the LReID-Hybrid stream Market-1501 $t$8 LTCC $t$9 MSMT17 $\mathcal D_{\rm train}^{(t)}$0 PRCC, Teata reports, at the last step, $\mathcal D_{\rm train}^{(t)}$1 on Market1501, $\mathcal D_{\rm train}^{(t)}$2 on LTCC, $\mathcal D_{\rm train}^{(t)}$3 on MSMT17, and $\mathcal D_{\rm train}^{(t)}$4 on PRCC in mAP/Rank-1; the seen-domain SC average is $\mathcal D_{\rm train}^{(t)}$5 and the CC average is $\mathcal D_{\rm train}^{(t)}$6. On unseen domains, it reports $\mathcal D_{\rm train}^{(t)}$7 on CUHK01, $\mathcal D_{\rm train}^{(t)}$8 on CUHK02, $\mathcal D_{\rm train}^{(t)}$9 on GRID, $\mathcal G_{t-1}$0 on SenseReID, $\mathcal G_{t-1}$1 on PRID, $\mathcal G_{t-1}$2 on VC-Clothes, and $\mathcal G_{t-1}$3 on Celeb-reID-light (Wang et al., 2024).

4. Subsequent frameworks and training paradigms

Later work broadens the design space rather than converging on a single solution. CMLReID is a CLIP-based framework with two modules: Context-Aware Semantic Prompt (CASP) and Adaptive Knowledge Fusion and Projection (AKFP). CASP forms contextualized text embeddings from a visual feature, a context encoder, base semantic prompts, and dynamic modulation prompts. AKFP maintains SC and CC prototypes, predicts clothing-state probabilities, and fuses two projection heads into

$\mathcal G_{t-1}$4

trained with a Clothing-State-Aware Projection Loss plus identity and triplet losses. The method does not store raw past images and uses prototypes as compact memory (Long et al., 14 Sep 2025).

DASK addresses exemplar-free lifelong ReID through Distribution Rehearser Learning and Distribution Rehearsing-driven LReID Training. A lightweight AKPNet predicts an instance-specific convolutional kernel $\mathcal G_{t-1}$5 from augmented current-domain images and learns to reconstruct the original image with an $\mathcal G_{t-1}$6 loss. At the next step, the old AKPNet generates old-style images

$\mathcal G_{t-1}$7

which are used jointly with current data under ReID and similarity-knowledge-distillation objectives. DASK reports that, relative to DKP, it improves Seen-Avg by $\mathcal G_{t-1}$8 to $\mathcal G_{t-1}$9 mAP and UnSeen-Avg by $\mathcal G_t$0 to $\mathcal G_t$1 mAP across two curriculum orders (Xu et al., 2024).

SPRED extends lifelong ReID into the semi-supervised regime. Its setting uses domains

$\mathcal G_t$2

with few labeled images and many unlabeled ones, no replay, and disjoint identities across tasks. The framework introduces learnable identity prototypes, neighbor-prototype pseudo-labeling, old-knowledge aligning–based clustering, new knowledge clustering, and a structure-preserving knowledge-distillation loss

$\mathcal G_t$3

At a $\mathcal G_t$4 label rate, SPRED reports Seen-Avg/UnSeen-Avg mAP of $\mathcal G_t$5, compared with $\mathcal G_t$6 for LSTKC and $\mathcal G_t$7 for HDC; SPRED$\mathcal G_t$8 reports $\mathcal G_t$9 mAP (Xu et al., 2 Jul 2025).

These systems differ in memory assumptions. Teata uses no data replay, CMLReID does not store raw past images, DASK stores no exemplars, and SPRED is also no-replay. The shared pattern is that forgetting is addressed through text prototypes, distribution rehearsal, or prototype evolution rather than historical image buffers.

5. Benchmarks, metrics, and empirical patterns

The benchmark introduced with Teata uses four sequential domains: Market-1501 (SC), LTCC (CC), MSMT17 (SC), and PRCC (CC), with four sequential orders; detailed results are reported for Order 1, Market-1501 $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$0 LTCC $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$1 MSMT17 $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$2 PRCC. Metrics are mAP and Rank-1, together with averages over seen SC domains $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$3, seen CC domains $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$4, and similarly for unseen domains (Wang et al., 2024).

A later formulation reports six different sequences of the four domains and evaluates after each new domain on all seen domains, with final SC-Average, CC-Average, and Total-Average (Long et al., 14 Sep 2025). Exemplar-free and semi-supervised extensions use different domain streams and therefore are not numerically interchangeable with the original hybrid benchmark, even when they are conceptually relevant (Xu et al., 2024, Xu et al., 2 Jul 2025).

Framework	Evaluation setting	Reported outcome
Teata	Order 1, last step	SC avg $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$5, CC avg $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$6
CMLReID	Order 1, seen domains	SC avg $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$7, CC avg $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$8, Total avg $\{\mathcal D_{\rm test}^{(t)}\}_{t=1}^T$9
DASK	Two curriculum orders, exemplar-free	Seen-Avg up to $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$0, UnSeen-Avg up to $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$1
SPRED	Semi-LReID, $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$2 label rate	Seen-Avg $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$3, UnSeen-Avg $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$4

Several empirical regularities recur. First, cloth-changing performance remains substantially below same-cloth performance. In Teata, seen-domain SC average is $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$5, whereas CC average is $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$6; in CMLReID, the corresponding figures are $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$7 and $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$8 (Wang et al., 2024, Long et al., 14 Sep 2025). Second, memory-free or replay-free learning is feasible but typically requires a stronger intermediate structure than plain feature distillation. Third, joint training is not a proxy for lifelong learning: on Teata’s Order 1 table, Joint-Train achieves $\{\mathcal D_{\rm test}^{(u)}\}_{u=1}^U$9 on SC average but only $t$0 on CC average, whereas Teata achieves $t$1 and $t$2 respectively (Wang et al., 2024). This suggests that simultaneous access to all data does not automatically resolve clothing-state mismatch.

6. Relation to multimodal and instruction-driven person retrieval

The hybrid-clothing-state literature sits within a broader shift toward multimodal person retrieval. Instruct-ReID defines a new task in which a query image or image-template is combined with an instruction that may be a short sentence, a full paragraph, or even another image. Existing six ReID tasks are treated as special cases by designing different instructions, and the baseline combines a frozen ViT image encoder, an instruction encoder, an Editing Transformer, and an adaptive triplet loss that modulates margin by instruction similarity. Its reported multi-task model improves mAP by $t$3, $t$4, and $t$5 on Market1501, MSMT17, and CUHK03 for traditional ReID; $t$6, $t$7, and $t$8 on PRCC, VC-Clothes, and LTCC for clothes-changing ReID; and $t$9 on COCAS+ real2 for language-instructed ReID (He et al., 2023).

MLLMReID adapts a pretrained multimodal LLM to ReID through a Common Instruction and a multi-task synchronization module called DirectReID. It uses a LLaMA-2-7B language core, a CLIP ViT-L/14 visual encoder, and a total loss

$$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$0

with $$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$1. On MSMT17 it reports $$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$2 mAP and $$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$3 Rank-1, compared with $$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$4 for its LLaVA+ReID baseline (Yang et al., 2024).

LLaVA-ReID introduces interactive person re-identification (Inter-ReID), where retrieval is refined over dialogue rounds. It uses a CLIP dual-stream retriever, a selective visual context module, and a LLaVA-OneVision-Qwen2-7B questioner trained with looking-forward question supervision. On Interactive-PEDES it reports $$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$5 R@1, $$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$6 R@5, $$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$7 R@10, $$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$8 mAP, and $$D_{t} = \{(I_{i}^{t}, y_{i}^{t}, s_{i}^{t})\}_{i=1}^{N_{t}},$$9 BRI, and it improves standard text-to-image ReID backbones such as IRRA and RDE after five rounds of question answering (Lu et al., 14 Apr 2025).

These systems do not define LReID-Hybrid in the narrow clothing-state sense, but they clarify why text space, prompt adaptation, and instruction conditioning recur in hybrid lifelong work. A plausible implication is that future LReID-Hybrid systems may further integrate continual learning with instruction-following, dialogue, or MLLM-based semantic compression.

7. Limitations and open directions

Reported limitations remain substantial. For Teata, limitations include dependency on the CLIP text encoder, the possibility that fixed prompt length $$\mathcal D^{(t)}=\{(x_i^{(t)},y_i^{(t)})\}_{i=1}^{N_{\rm img}^{(t)}}$$00 may not capture arbitrary attribute cardinalities, and the risk that the slow-paced learner may still cause slow drift of old semantics in extremely long task sequences (Wang et al., 2024). For CMLReID, CC performance remains significantly lower than SC performance, reliance on an explicit SC/CC binary classifier may struggle with subtle “partial” clothing changes, and novel domain shifts far away from seen clothing styles may still degrade performance (Long et al., 14 Sep 2025). For SPRED, distribution alignment assumes color-style shift, prototypes may overfit scarce labels at extremely low label rates, and memory-free clustering adds overhead (Xu et al., 2 Jul 2025).

Future directions already proposed in the literature include dynamically growing or pruning prompt slots, multilingual or graph-structured semantic slots, text-to-image generative regularization, multi-level clothing-change quantification beyond binary SC/CC, and lightweight exemplar replay or generative rehearsal (Wang et al., 2024, Long et al., 14 Sep 2025). Taken together, these proposals indicate that LReID-Hybrid is evolving from a narrowly defined continual benchmark into a broader research program on stable identity representation under sequential domain shift, clothing variation, and restricted memory.