Automatic Synthetic Data and Fine-grained Adaptive Feature Alignment for Composed Person Retrieval

Thank you very much for your valuable feedback and positive assessment. We reply to each of your concerns below.

Responses to Weaknesses/Questions:

A1.1: Efficiency and Computational Cost of Using Multiple Feature Vectors

During inference, existing Composed Image Retrieval (CIR) methods vary significantly in terms of the number of target-image tokens they use. For instance, among methods based on BLIP2 (CaLa and SPRC), CaLa only uses the first token as a global image feature, while SPRC uses all 32 tokens and selects the one with the highest similarity score as the final similarity. Regardless, the primary computational cost arises from feature extraction rather than feature storage or retrieval operations.

In terms of storage, CaLa stores a single 768-dimensional vector per image, while both SPRC and our FAFA store 32 vectors (each with 256 dimensions). For a database containing 10,000 images stored in half-precision format, CaLa requires approximately 15.36 MB, while SPRC and FAFA require about 163.84 MB. Such an amount of storage is relatively easy to manage.

We evaluate the inference efficiency on the entire ITCPR dataset using a single H800 GPU (the average of 10 runs, with the time unit being seconds):

These results show that feature extraction dominates computational cost. The retrieval step itself (involving simple matrix operations) consumes minimal time. Specifically, our FAFA achieves the fastest query feature extraction as it involves only a Q-Former and a visual encoder, unlike SPRC and CaLa, which include additional text encoders. Although CaLa can extract candidate features more quickly, its retrieval speed is the slowest. Importantly, varying the parameter k in FAFA (the number of tokens used for similarity computation) has a negligible impact on retrieval time. Therefore, selecting k based on retrieval accuracy is optimal; the original Figure 5(b) reports retrieval performance for k values from 1 to 16, with k = 6 achieving the best performance and thus chosen as our final setting.

A1.2: Practical Justification for the CPR Task

Thank you for raising this practical question. The Composed Person Retrieval (CPR) task has significant application value in multiple aspects:

• Missing person search: Families usually provide a past photograph and verbal descriptions of the current appearance of the person involved. A CPR system integrating both image and textual descriptions can efficiently give the most relevant results from surveillance or social media images, etc. This situation happens frequently, especially with children and the elderly.
• Media editing and content management: Video editors can efficiently retrieve specific clips using actor photographs combined with descriptions of desired scenes or states.
• Personalized fashion recommendation: Users on e-commerce platforms can upload personal photos along with textual style descriptions, thus precisely retrieving matching outfits and enhancing shopping experiences.

Therefore, CPR has broad applicability. We will incorporate this discussion in the revised manuscript.

A1.3: Intuitions and Analysis of Feature Diversity (FD) and Masked Feature Reasoning (MFR)

Feature Diversity (FD)

Intuition: In the CPR task, fine-grained target features precisely capture subtle appearance differences associated with the compositional query. Without FD supervision, extracted features may become redundant, limiting the ability to distinguish delicate differences. FD supervision encourages the extracted features to be sufficiently distinctive and complementary, thus fully representing the rich details within person images.

Qualitative Analysis: FD loss sets a maximum cosine similarity threshold m, pushing features apart in the semantic space. This helps form diverse representations and enhances sensitivity to subtle variations. Specifically, the model trained with FD (No.5 in Table 2) reduces the intra-group maximum cosine similarity of target features from 0.824 (No.4 in Table 2) to 0.713, indicating that it contains richer and more diverse information and is suitable for Fine-grained Dynamic Alignment (FDA).

Theoretical Analysis: From the perspective of representation learning, FD acts similarly to a regularization mechanism, maximizing the semantic coverage and discriminability of the target feature space. From an information-theoretic viewpoint, FD reduces redundancy and uncertainty and increases entropy, thus enhancing representational efficiency and generalization, especially when the test distribution is unknown.

Masked Feature Reasoning (MFR)

Intuition: In the CPR task, the combination of the reference image and textual description inherently involves complementary information, as the text describes visual changes. MFR randomly masks parts of the combined representation to force the model to explicitly capture this complementarity, thus improving its capability to recover complete representations from partial information.

Qualitative Analysis: MFR enhances the extraction of complementary information between reference images and text, thus improving alignment with target features. An intuitive demonstration method is to visualize the attention distributions when extracting query features and observe if textual emphasis and image regions become more complementary after adding MFR supervision. We will include this visualization in the revised manuscript (images cannot be inserted in the current reply). Additionally, the average cosine similarity between query features and FDA-derived target features increases from 0.481 (No.3 in Table 2) to 0.504 (No.5 in Table 2), indirectly confirming MFR's effectiveness in promoting complementary information extraction.

Theoretical Analysis: Theoretically, MFR is similar to masked autoencoder mechanisms, improving the model's understanding of input features through a self-supervised masking constraint. Randomly masking feature dimensions can act as a self-supervised objective, promoting effective modeling of conditional probability distributions between different modalities. This implicit reconstruction objective increases the density and continuity of the cross-modal feature space and theoretically ensures that more robust and generalizable representations can be obtained.

Response to Limitations:

A2: Thanks for your forward-looking comments. A more detailed discussion regarding the limitations will be expanded in the revised version. Firstly, following mainstream person retrieval approaches, FAFA currently only supports taking one image and one textual description as input. However, a feasible approach for introducing multiple reference images or textual descriptions to further improve practicality is discussed in our response to Reviewer iMvu (A2.3). Additionally, the application of the CPR model under resource-constrained conditions also remains to be further addressed.

Response to Ethical Concerns:

A3.1: Data Quality and Representativeness

Our SynCPR dataset is entirely generated through an automated pipeline, involving three stages: the generation of diverse textual quadruples, the synthesis of identity-consistent images, and rigorous multimodal quality filtering. Specifically, approximately 140,000 diverse textual quadruples are generated by Qwen2.5-70B; subsequently, identity-consistent image pairs are synthesized by the Flux.1 model fine-tuned via LoRA, with additional style perturbations to enhance diversity. Finally, a strict multi-criteria filtering approach retains 1.15 million high-quality triplets. These triplets cover extensive variations in clothing, age, scene, and so on, ensuring high diversity and representativeness (Supplementary Figure S10 provides an intuitive illustration of this point). Therefore, there are no risks arising from biases or distortions.

A3.2: Human Rights and Surveillance Concerns

It should be noted that: (1) Our SynCPR dataset is completely synthetic; all person images in it depict entirely fictional individuals, without involving any real identities or non-public materials. (2) Our ITCPR test set is constructed from publicly available datasets (Celeb-reID, PRCC, LAST), explicitly permitting secondary editing and academic usage. Therefore, there are no risks in terms of ethics or human rights.

Thank you very much for your valuable and insightful feedback. We carefully address each of your concerns as follows.

Responses to Weaknesses:

A1.1: Insufficient Experimental Validation

Considering the space limitations, we arrange more experimental details and supplementary results in the Appendix. It is worth pointing out that Table 1 is not just a performance comparison, but a comprehensive summary that addresses your multiple key concerns. Firstly, regarding the rationality and uniqueness of the new Composed Person Retrieval (CPR) task, the Introduction highlights its practical significance: in real-world scenarios, visual and textual information are often both accessible. For example, when elderly individuals or children go missing, which occurs frequently, we usually have previous photos as well as verbal descriptions of their current appearance. Therefore, CPR is specifically designed to leverage such complementary multimodal information for person retrieval in realistic search scenarios. Please refer to the response for Reviewer AeXC (A1.2) for more potential application examples.

Secondly, regarding empirical validation, the first three experimental settings in Table 1 demonstrate that relying solely on either visual or textual modalities (as is common in existing person retrieval methods) is ineffective for CPR. In contrast, a simple combination of image and text (the 4th setting in Table 1) substantially improves retrieval performance, validating both our CPR task formulation and the relevance of the ITCPR annotations. Further improvements through our proposed FAFA framework and training on SynCPR show the importance of effectively extracting complementary multimodal information. These results also indicate that the proposed modules are indispensable.

We will provide more detailed analysis and interpretation in the revised manuscript. Additionally, we have added new experiments to strengthen the sufficiency. Please refer to the responses for Reviewer iMvu (A1.1 and A2.2), Reviewer ZBdY (A3.1), and Reviewer AeXC (A1.1).

A1.2: Privacy Concerns and Negative Social Impacts

Thank you for raising these ethical concerns. It should be noted that: (1) Our SynCPR dataset is completely synthetic, with all person images depicting entirely fictional individuals, without involving any real identities or non-public materials. (2) Our ITCPR test set is constructed from publicly available datasets (Celeb-reID, PRCC, LAST), explicitly permitting secondary editing and academic usage. Therefore, there are no risks in terms of privacy issues or negative social impacts. In the revised version, we will modify and clarify the descriptions that may lead to potential risks, to reflect the social responsibility and core value of "Tech for Good".

A1.3: Insufficient Comparisons with Related Works

Thank you for suggesting relevant works, and we will add discussions of these methods in our revised Related Work section. However, it should be noted that these methods are not quite suitable for the CPR task we propose. For instance, the InstructReID model typically uses fixed textual instructions for most individuals. When it receives flexible textual instructions combined with images, the textual content tends to duplicate rather than complement the image information. Essentially, this belongs to modality fusion rather than complementary multimodal input. Consequently, InstructReID is not suitable for CPR scenarios that require complementary rather than redundant information.

Moreover, the UNIReID model is limited to forensic scenarios involving sketch-plus-text retrieval and has not publicly released its trained checkpoints, limiting direct comparisons. We thus focus on experimentally comparing InstructReID's inference performance on our CPR task:

Clearly, the InstructReID model fails to effectively utilize complementary multimodal information for CPR. The performance of its combined input is even lower than that of single-modal textual input, highlighting FAFA's superiority for this specific task.

Responses to Questions:

A2.1: Identity Consistency in Data Generation (Quantitative Validation)

As described in Section 3.1, our data synthesis pipeline ensures identity consistency by generating two related sub-images simultaneously within a single image. Qualitative evidence is provided by Figure 3 in the manuscript and Figure S10 in the Appendix. Moreover, we agree that quantitative verification is necessary, and thus we conduct an additional experiment. Specifically, we randomly select 10 000 triplets (5 000 quadruples) from SynCPR and compare two generation strategies—individual image generation and simultaneous paired generation—using identical prompts and the same fine-tuned Flux.1 model. To ensure a fair comparison, generated images that do not meet the multimodal filtering criteria (excluding identity consistency) are regenerated.

We utilize the FIRe2 [1] model (pre-trained on LTCC [2]) to compute an Identity Recall metric. For each of the 10 000 images (5 000 identities), we retrieve the top-10 matches from the remaining images; if the same identity appears in these matches, it is regarded as a successful recall. Identity Recall is the percentage of successful recalls. Retrieval performance (Rank-1, Rank-5, Rank-10, mAP) using FAFA trained on these datasets is also reported:

These results clearly demonstrate the effectiveness of our simultaneous-generation approach in ensuring identity consistency. This quantitative analysis will be included in the revised manuscript.

[1] Exploring fine-grained representation and recomposition for cloth-changing person re-identification. TIFS, 2024

[2] Long-term cloth-changing person re-identification. ACCV, 2020

A2.2: Choice of Dataset for Fine-tuning

Regarding the concern about whether all datasets should be used: our goal is to propose a feasible solution for the new CPR task, which serves as an important supplement to existing person-retrieval technologies such as text-based person retrieval (TPR), image-based person retrieval (IPR), and cloth-changing Re-ID. Therefore, UFine6926, which primarily collects images from online videos, is excluded. Among the three classic TPR datasets (CUHK-PEDES, ICFG-PEDES, RSTPReid), only CUHK-PEDES gathers images from five surveillance datasets, providing broader scene diversity. The other two datasets source their images from MSMT17, which mainly contains a single parking-lot scenario. To prevent overly homogeneous scene distributions from negatively affecting generation diversity, we choose CUHK-PEDES for fine-tuning. We will briefly add this reasoning to the revised manuscript.

A2.3: Fine-tuning TPR Methods

We clarify a misunderstanding: all compared methods, including our own, are independently trained without referencing the ITCPR test-set distribution, meaning that every experiment constitutes a cross-domain test. Following your suggestion, we fine-tune two representative TPR methods (IRRA and RDE) on our SynCPR dataset for a fair comparison. Reference images and textual descriptions are combined into a global query feature by summing features from dual encoders, while target images are encoded into global features separately. We train these methods under two settings: training from scratch and fine-tuning from pre-trained checkpoints on TPR datasets. The results are as follows:

These results show that pretraining on TPR datasets improves retrieval performance. Nevertheless, TPR methods remain far inferior to our FAFA and the fine-tuned Composed Image Retrieval (CIR) approaches (Table 1). CIR methods leverage attention-based interactions between reference images and relative textual descriptions, whereas TPR methods encode these independently, limiting their ability to exploit complementary multimodal information effectively.

A2.4: Cross-domain and Cross-dataset Experiments

As explained in A2.3, all of our current experiments are conducted under cross-domain settings. To the best of our knowledge, CPR is a newly proposed task, and there are presently no alternative datasets available for broader testing. Nevertheless, in Table S3 of the Appendix we additionally compare FAFA on a CIR benchmark that also uses triplet supervision; FAFA still achieves state-of-the-art performance, further demonstrating the effectiveness and robustness of our method. We welcome the Reviewer to provide datasets suitable for cross-dataset validation for the additional experiments.

Response to Limitations:

A3: We will significantly expand our discussion of limitations and potential negative social impacts as described in response A1.2. Additionally, introducing multiple reference images or textual descriptions to improve practicality, as well as deploying the CPR model under resource-constrained conditions, remains to be explored.

We sincerely thank you for your efforts and time, as well as your overall positive evaluation. In response to your crucial comments, we provide detailed clarifications below.

Responses to Weaknesses:

A1.1: Focus Mainly on Appearance Variations

When manually annotating the ITCPR test set, we primarily focus on appearance changes. This focus reflects many practical scenarios. For instance, when looking for a missing person, we may have a previous photograph in which the appearance often differs significantly from that on the day of disappearance. Considering that other types of variations, as you pointed out, such as changes in scenes or hairstyles and so on, also hold additional practical value, we plan to add more annotations to align with our SynCPR training set, which includes various types of variations.

A1.2: Figure Icon Issues

Thank you for pointing out the issue with the icons in Figures 1(b) and 3. In the revised manuscript, we will refine both figures by adding clearer legends and supplementary annotations to improve readability and ensure that all icons are easily distinguishable.

A1.3: This concern overlaps with our response provided in A1.1. Please refer to A1.1.

A1.4: Distinction Between CPR and CIR Tasks

Composed Image Retrieval (CIR) aims to retrieve a target image based on a query composed of a reference image and a relative caption that describes the difference between the two images [1]. Composed Person Retrieval (CPR) aims to retrieve a target person image. Therefore, the main distinction between CPR and CIR lies in the granularity and strictness of retrieval constraints. CIR focuses on retrieving images matching the semantic modifications described by a combination of textual and visual queries, and there are no strong constraints between the reference and target images. Typical CIR examples are provided by Figure 8 in the arXiv version of reference [1]. In contrast, CPR requires the retrieved target person to share the SAME ID as the reference individual, supplemented with additional details such as clothing attributes, thereby imposing stronger constraints. Moreover, since all images in CPR are of persons, the retrieval model must precisely distinguish subtle semantic differences among numerous visually similar candidates. Therefore, compared with CIR, the CPR task requires stronger capabilities in image relevance assessment and fine-grained semantic differentiation.

[1] Zero-Shot Composed Image Retrieval with Textual Inversion. ICCV2023 or arXiv:2303.15247

Responses to Questions:

A2.1: Clarification on Using Multiple β Values

When the data distribution of the test domain is unknown, increasing the diversity of training data helps to enhance the model's adaptability under various distributions. To this end, via LLM-generated texts, our method first ensures diverse textual descriptions. Then, the image style diversity is controlled by adjusting the β value. As shown in supplementary Figure S11, different β values correspond to different image styles. For example, β=0 corresponds to purely generated (artificial) style images; β≈0.3 corresponds to photographic-quality images; β=1 corresponds closely to surveillance-style images. Therefore, to achieve diverse image styles and maintain a strong sense of realism for the majority of images, we adopt a strategy of setting β=1 with 50% probability during inference, while randomly sampling β values in other cases to ensure broad distribution coverage.

To verify the impact of different β strategies on the retrieval performance of FAFA, we first randomly select 10,000 triplets (5,000 quadruples) from the SynCPR dataset as a benchmark training set. Then, we respectively generate different datasets using fixed β=1, β=0, and fully random β. To ensure a fair comparison, the images that do not meet the filtering criteria are regenerated. The performance comparison is as follows:

The results clearly show that predominantly using unrealistic images (β=0) seriously harms retrieval performance. Although a fully random β can provide broad distribution coverage, a high proportion of less realistic images slightly reduces retrieval effectiveness. Conversely, relying exclusively on highly realistic images (β=1) narrows the data distribution, causing a slight performance drop due to varying image styles from multiple sources within ITCPR. Finally, the strategy of combining realistic images with broad distributional coverage achieves the best performance. We will add the above analysis to the revised manuscript.

A2.2: Reliability of FAFA Superiority Claims in Table 1

In Table 1, we aim to compare all promising solutions for the CPR task. To objectively evaluate the performance of each method on the CPR task, the selected training datasets for these methods can be either the ones recommended in their original works or the widely recognized datasets within their respective fields. We observe that except for the first two experimental settings (which are not specifically trained for text+image retrieval) failing to reliably demonstrate FAFA's superiority, the comprehensive results under all other settings effectively validate the superior performance of FAFA.

It should be noted that the first two settings are actually essential for demonstrating the validity of the ITCPR dataset construction and related experimental protocols. In other words, using only text, only images, or simply combining the two leads to poor performance. Only by effectively extracting complementary information from combined queries, as demonstrated by FAFA, can higher retrieval accuracy be achieved.

A2.3: Fine-grained Features Derived from Target Image The reasons for extracting fine-grained features solely from the target image are as follows:

1. Retrieval efficiency: If fine-grained features are used on both query and target sides, matching will require solving a (k × k) bipartite graph, significantly increasing computational complexity and complicating loss design. Additionally, this approach violates our "dynamic alignment" principle, which emphasizes "selection on one side and aggregation on the other."

2. Clear role differentiation aligned with CPR characteristics:

   • Query side: Combines a reference image and relative textual description, inherently requiring information integration to predict "how the appearance should change."
   • Target side: Must identify the correct target from numerous candidates, inherently requiring preservation of local discriminative features for fine-grained differentiation.

3. Avoiding multimodal noise interference: On the query side, the image and textual description are integrated through the Q-Former, resulting in a semantically-completed holistic representation. The Masked Feature Reasoning (MFR) further reinforces this integration. If this unified representation is decomposed back into patch-level fine-grained vectors, textual noise and semantic confusion will be introduced at the patch level, negatively affecting retrieval performance.

Responses to Limitations:

A3.1: Modality Deficiency Experiments

Thank you for providing such an insightful suggestion. Our FAFA can handle the situation of missing modalities by making minor adjustments to the training strategy. Specifically, while keeping all other settings unchanged, we retrain FAFA on SynCPR data with several modality-deficient scenarios: 70% complete inputs, 15% missing textual inputs (blank text), and 15% missing images (purely black images). During inference, the same modality-deficiency settings are applied. The resulting test performances are as follows:

Comparing these results to those reported in Table 1, we can see that the modified FAFA slightly reduces performance under combined inputs but remains highly effective. Moreover, FAFA performance under modality-deficient conditions surpasses many dedicated image-based person retrieval (IPR) [2,3] and text-based person retrieval (TPR) methods [4,5], which are specifically trained on their respective benchmark datasets.

[2] Beyond appearance: a semantic controllable self-supervised learning framework for human-centric visual tasks. CVPR, 2023

[3] Exploring fine-grained representation and recomposition for cloth-changing person re-identification. TIFS, 2024

[4] Noisy-correspondence learning for text-to-image person re-identification. CVPR, 2024

[5] RaSa: relation and sensitivity aware representation learning for text-based person search. IJCAI, 2023

A3.2: Please see the response to A1.1 above.

We appreciate your efforts and time, as well as your overall positive assessment. In response to your crucial concerns, detailed clarifications are provided below.

Response to Weaknesses:

A1.1: In-depth analysis of the domain gap between synthetic training data and real-world test data

Thank you for this insightful suggestion. For the Composed Person Retrieval (CPR) task, the textual data primarily describes relative appearance differences between the target and reference person images. Therefore, the main source of potential domain discrepancy originates from differences in visual distributions between synthetic and real-world images.

During the construction of the SynCPR dataset, apart from knowing that the test images are real images, we have no prior access to any distributional information about the real-world test images. Consequently, we aim to ensure that the majority of images in the SynCPR are highly realistic and sufficiently diverse in distribution. Specifically, the image generator used in the SynCPR dataset is fine-tuned on the real-world surveillance dataset CUHK-PEDES. During inference, we adopt a mixed strategy: setting β=1 with a 50% probability to ensure highly realistic images, and randomly setting β with a 50% probability to encourage broad distributional diversity (partially overlapping with the real-world test domain). Its purpose is to minimize distributional differences in the absence of explicit test-domain information.

To quantitatively assess this strategy, we calculate the Fréchet Inception Distance (FID) using different settings. Specifically, we randomly sample 5,000 image groups (10,000 images in total) from SynCPR and compute the FID against all images annotated in the real-world ITCPR test set, using the same prompts but different β settings. The results are as follows:

These results demonstrate that our current strategy effectively minimizes the domain gap between synthetic training images and real-world test images.

Theoretical justification: Generally, when two distributions are narrow with minimal overlap, the FID value is relatively large. Conversely, when one distribution is broad (representing diverse SynCPR data) and the other is narrow (representing ITCPR data), with some overlap, the FID tends to be lower [1, 2].

[1] GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. NIPS, 2017 [2] Training Generative Adversarial Networks with Limited Data. NIPS, 2020

A1.5: Improvements of retrieval framework

In the field of person retrieval, especially for the text-to-image retrieval task, it is a common practice to further fine-tune pre-trained Vision-Language Pretraining (VLP) models. Leveraging pre-trained knowledge enables rapid and effective alignment between textual descriptions and visual content, thereby achieving superior performance. Given that the CPR task also inherently involves cross-modal retrieval, we therefore select the visual encoder and Q-Former components of BLIP-2 as the foundational architecture. Importantly, we further propose three new modules to specifically address the fine-grained nature of person retrieval: Fine-grained Dynamic Alignment (FDA), adaptive Feature Diversity (FD) supervision, and bidirectional Masked Feature Reasoning (MFR), aiming to make full use of complementary information between reference images and relative captions. Comprehensive experiments and ablation studies verify the positive effect of each component on performance improvement. Although these improvements only constitute one aspect of our overall contributions, they have shown indispensable value in the context of our proposed retrieval framework.

A1.6: Scale of test dataset

In the Text-based Person Retrieval (TPR) field, there are two widely-used benchmark datasets, namely the CUHK-PEDES and the RSTPReid. The former contains 3,074 images and 6,156 texts in its test set, while the latter only includes 1,000 images and 2,000 texts for testing. In contrast, our ITCPR test set contains 2,225 triplets (reference image, relative caption, target image) and requires more complex and meticulous annotations; thus, its scale is comparable to that of similar datasets and is capable of supporting the validation of model effectiveness. Considering that this is the first dataset in the field of CPR, the ITCPR can be expanded in terms of both scale and quality. Hence, we will follow your comment to continue with the annotation and expansion work in our subsequent steps while retaining the current version (so that other methods can conduct fair comparisons).

Responses to Questions:

A2.1 Computational Efficiency of Data Synthesis

We conduct efficiency evaluations on a single H800 GPU following the configurations described in Section 4.1. The results show that: (1) With batch sizes set to 8, the average inference time for generating a single textual quadruple using Qwen2.5-72B is approximately 6.45 seconds. (2) The generation of a pair of person images takes around 3.23 seconds. (3) Filtering a generated image pair using Qwen2.5-VL-32B requires about 0.38 seconds. Therefore, synthesizing one annotated quadruplet, which consists of two triplet annotations, takes an average of 10.06 seconds, with each individual triplet annotation requiring 5.03 seconds. Consequently, approximately 17,000 annotated triplets can be generated daily on a single GPU, which is approximately 24 times faster than manual annotation (it takes an average of 120 seconds for one triplet), making our pipeline highly practical for large-scale applications. It is worth mentioning that this synthesis process only needs to be carried out once and does not require repeated computation. Furthermore, the SynCPR dataset will certainly be open-sourced, thus avoiding the need for redundant generation efforts in the future.

A2.2 Robustness of FAFA Framework

As detailed in Appendix C.1, we incorporate strong data augmentation techniques during FAFA training, including random horizontal flipping, random cropping and padding, and random erasing. These augmentations enhance the robustness of the model to severe occlusions and low-quality reference images. Specifically, we evaluate the FAFA model trained on the SynCPR dataset under various perturbations applied to the ITCPR dataset. These perturbations include random erasing of image regions (covering 5%-30% of the image area) to simulate severe occlusions, and applying Gaussian blurring with randomly chosen standard deviations (with σ from 0.1 to 3.0) to simulate low-quality images. The results are summarized below:

The results indicate that our FAFA demonstrates strong robustness. It can maintain its effectiveness even under severe occlusion and low-quality conditions, and the performance degradation is within a controllable range. Additionally, in response to Reviewer ZBdY's comment A3.1, we further discuss the performance of FAFA when encountering modality deficiencies, and the favorable results also support the robustness of our method.

A2.3 Feasibility of Multi-image and Multi-text Input

Thank you for keenly identifying this promising direction for future research, as we envisioned in the conclusion. However, it is difficult to provide quantitative results in a short time because the corresponding test sets require high-quality manual annotations, and there is currently a lack of publicly available datasets that support multiple reference images and textual descriptions.

Nevertheless, from the perspectives of dataset synthesis and methodological extensions, we have designed a feasible technical solution. Firstly, for data generation, the established SynCPR dataset can be used as training data for image generators, enabling the training of consistent generation or image-editing models [3,4]. Such methods are helpful for generating multiple diverse images of the same individual based on one reference image combined with varied textual descriptions. Then, the textual generation and filtering pipeline proposed in our paper can naturally extend to construct a dataset supporting multiple references for the CPR task.

Regarding the training of the retrieval model, simple modifications can enable it to adapt to the situation of multiple images or texts. One straightforward approach is to increase the dimensionality of image input and use placeholders when necessary. Meanwhile, multiple textual descriptions can be naturally concatenated into a single, extended query sentence.

[3] Ominicontrol: Minimal and universal control for diffusion transformer. ICCV, 2025

[4] In-Context Edit: Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer. arXiv:2504.20690

Reviewer iMvu, please engage in the discussion period. I understand that the review period started over a weekend, but we only have a few days remaining in the (now slightly extended) discussion period. The authors have provided a thoughtful response to your review and you are obligated to respond to it. You should share with the authors if they addressed questions or concerns you had, and seek clarification about any questions or concerns that remain. Please post your response as soon as you can so that there is time for the authors to follow up and discussion to progress as needed.

Thanks for your efforts in addressing my concerns. I believe the issues have been mostly well explained, and I will maintain my positive score.