acceptodds
Under review as a conference paper at ICLR 2027

Selective Attention for Training-Free Consistent Text-to-Image Generation

Abstract

Generating images with consistent object appearances across different scenes remains a fundamental challenge in text-to-image diffusion models. Existing approaches either require expensive fine-tuning for each subject or suffer from feature leakage when unconditionally sharing attention across images, where irrelevant backgrounds contaminate the target object's appearance. To address this, we first analyze the root cause of background leakage in naive cross-image self-attention fusion, identifying softmax normalization as the key factor that amplifies irrelevant token contributions. Grounded in this analysis, we propose SelectDiff, a training-free framework based on similarity-guided selective attention. Instead of fusing all cross-image tokens, we retain only the top-K most relevant connections per query token, naturally filtering background contamination while preserving subject identity. We further propose Next-Step Variance Ascent (NSVA), which dynamically adjusts K across denoising steps by maximizing between-class attention variance, and foreground-background separation to restrict cross-image attention exclusively to subject-relevant tokens. Experiments on the ConsiStory+ benchmark demonstrate that SelectDiff outperforms both fine-tuning-based and training-free methods in identity preservation and prompt alignment, while achieving superior inference efficiency. As a natural byproduct, our approach handles multi-subject scenarios without any specialized processing.

Primary area:
generative models
Keywords:
training-free consistent image generation, diffusion models, subject consistency

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_JE2vOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SelectDiff, a training-free method for subject-consistent text-to-image generation. Instead of sharing all cross-image attention tokens, each query attends only to its top-K most similar tokens in the other images. NSVA adapts K across denoising steps with an Otsu-style variance criterion, and FBS restricts cross-image attention to likely foreground queries. On ConsiStory+ with SDXL, the method reports better consistency, prompt alignment and speed than prior training-free approaches, and it extends naturally to multi-subject scenes.

Critical strengths
  1. It is a simple and elegant idea. Using query–key similarity itself as an implicit subject mask is a neat insight. It removes ConsiStory's explicit mask extraction while being plug-and-play and faster. The adaptive-K view based on between-class variance is also a creative angle.
  2. The method gives the best training-free CLIP-T, CLIP-I and DreamSim scores on SDXL. It handles multiple subjects without any special design (Fig. 4), and the authors honestly report a failure case.
Critical weaknesses
  1. Recent training-free methods, especially DiT-based ones (e.g. CharaConsist), are missing. The FLUX.2 row is compared only against SDXL baselines and is described as "reserved".
  2. NSVA is never compared quantitatively with the best fixed K or with the exact Otsu solution, and Fig. 6 is a single qualitative prompt. Table 2 includes an "FBS without SGSA" row, although FBS is defined through top-K scores.
  3. Key hyperparameters (η, Q-rate, affected layers and timesteps, sampler settings) are missing.
Questions for authors
  1. How is FBS implemented without SGSA (Table 2)?
  2. Could you add a quantitative comparison of NSVA vs. the best fixed K vs. exact Otsu, and a ConsiStory-style baseline on the same FLUX.2 backbone?
3 comments
AuthorsOpenReview ↗

Response to Reviewer JE2v

W1 and Q2. Recent DiT methods and the “reserved” wording

Thank you for this suggestion. We have added CharaConsist, a recent DiT-based method, and broadened the evaluation with aesthetic and VLM-based scores. Aesthetic quality is evaluated using LAION Aesthetic Predictor V2, with raw prediction scores reported. GPT-6 provides a custom composite assessment of text–image semantic alignment, visual quality, cross-image subject consistency, and background leakage. Higher scores indicate better overall performance across these dimensions, rather than an independent improvement in every component. We have included the evaluation prompts, model version, scoring scale, and aggregation procedure in the revised manuscript.

Table R1. Comparison with a recent DiT-based method.

MethodBackboneCLIP-T ↑CLIP-I ↑DreamSim ↓Aesthetic Score (V2) ↑VLM Score (GPT-6) ↑Time (s) ↓Memory (GB) ↓
CharaConsistDiT90.591.822.56.280.629.744.9
SelectDiff+ (Ours)DiT92.194.713.27.588.324.630.2

SelectDiff+ improves CLIP-T and CLIP-I by 1.6 and 2.9, reduces DreamSim by 9.3, and obtains higher aesthetic and VLM scores. Runtime decreases by 5.1 seconds and memory by 14.7 GB. These results extend the system-level comparison to a recent DiT-based consistency method, and we have incorporated them into the revised manuscript. Sharing the DiT architecture category does not by itself establish identical backbone checkpoints.

We also apologize for the retained “reserved” wording. The experiments were completed, but the placeholder was inadvertently left unchanged during manuscript preparation. We have replaced it with: “The SelectDiff (FLUX.2) row in Table 1 reports results obtained under the ConsiStory+ protocol.”

W2 and Q2. NSVA versus the best fixed K and exact Otsu

Thank you for recommending these comparisons. Exact maximization of the Otsu-style between-class variance is possible for a given set of attention scores. However, this is a single-step optimum, not a global optimum for the denoising trajectory or final generation quality, because the score distribution changes over time.

NSVA reduces the search overhead of adapting K: it evaluates the objectives at K and K+1, forms a local finite-difference update, and applies the update at the next denoising step. Exact maximization instead evaluates all candidate K values. After sorting, prefix sums allow this evaluation in O(L) additional work per score vector, whereas NSVA requires two objective evaluations once the necessary statistics are available. These counts exclude similarity computation and acquisition of the statistics; the following experiment reports generation quality and measured runtime.

Table R2. K-selection comparison using the same cache settings.

K-selection policyCLIP-T ↑CLIP-I ↑DreamSim ↓Time (s) ↓
Validation-selected fixed K85.387.622.417.2
Exact Otsu90.192.319.636.6
NSVA90.292.718.519.5

Compared with the validation-selected fixed K, NSVA increases CLIP-T and CLIP-I by 4.9 and 5.1, reduces DreamSim by 3.9, and increases runtime by 2.3 seconds. Compared with exact Otsu, its three quality metrics are numerically better, while runtime is reduced by approximately 46.7%. These results support its quality–runtime trade-off; better final quality does not imply more accurate optimization of the single-step Otsu objective. Fig. 6 illustrates parameter effects and is not used as evidence that NSVA outperforms every fixed-K setting.

W3. Missing hyperparameters and implementation details

Thank you for pointing out the missing implementation details. The configuration described here uses η=1, a Q-rate of 0.8, application to all self-attention layers, 20 denoising steps, and the DDIM sampler. We have provided the complete settings and their backbone-specific scope, including the affected denoising timesteps, to improve reproducibility.

Q1. How is FBS implemented without SGSA in Table 2?

Thank you for identifying this ambiguity. As described in §4.3 on page 8, removing SGSA replaces sparse aggregation with naive cross-image attention fusion. In the FBS-only configuration, we still compute foreground scores from the mean top-K similarities and select the top-Q queries. These selected queries then attend to all reference keys and values, without restricting cross-image connections to the top-K subset. Unselected queries use within-image self-attention only.

Thus, this configuration retains FBS foreground-query selection while removing sparse cross-image feature aggregation. We have clarified this implementation in §4.3 and explicitly distinguished the top-K operation used for foreground scoring from the top-K restriction used for feature aggregation.

Thank you to the authors for the detailed response and for providing the additional experimental results and implementation details. Overall, I find the rebuttal satisfactory and will update my score to 3.

Thank you for your careful review and for updating your score. We truly appreciate your constructive feedback, which has helped us substantially improve the paper.

Reviewer_TMb9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SelectDiff, a training-free framework for subject-consistent text-to-image generation. The method introduces similarity-guided selective attention (SGSA) to retain the top-K most relevant cross-image token connections, Next-Step Variance Ascent (NSVA) to dynamically adjust K during denoising, and foreground-background separation (FBS) to suppress irrelevant feature sharing. Experiments on the ConsiStory+ benchmark demonstrate improvements in subject consistency, prompt alignment, and inference efficiency.

Critical strengths

The paper addresses a practical issue in training-free consistent image generation, particularly background leakage caused by unrestricted cross-image attention. Selectively restricting feature sharing is intuitive and computationally appealing. The reported results show a reasonable balance between identity preservation, prompt alignment, and inference efficiency.

Critical weaknesses
  1. Limited methodological novelty. The proposed framework primarily combines existing techniques, including top-K sparse attention, variance-based threshold selection, and foreground-background filtering. In particular, NSVA is based on the classical Otsu between-class variance criterion, while FBS simply ranks query tokens according to their similarity scores. These components appear to be heuristic adaptations of established methods rather than fundamentally new attention mechanisms. The paper should clarify what constitutes its core technical innovation beyond combining these existing operations.

  2. Missing recent and closely related baselines. The comparison in Table 1 mainly includes methods from 2023–2025, while several highly relevant recent approaches are omitted, including CoDi (ICLR 2026), CharaConsist (ICCV 2025), ConsiStyle (SIGGRAPH Asia 2025), and SDeC (ICLR 2026). These methods directly address training-free subject consistency, foreground-background disentanglement, identity preservation, or pose and style diversity, and have publicly available implementations. Particularly, CoDi and CharaConsist provide important comparisons for identity-aware feature interaction. Without evaluating against these stronger and more recent approaches under comparable settings, the claimed superiority over existing methods is not sufficiently convincing.

  3. Insufficient justification of similarity-based subject selection. The method assumes that semantically corresponding subject tokens naturally exhibit higher similarity than irrelevant background tokens. However, high feature similarity does not necessarily imply instance-level identity correspondence, especially when multiple subjects belong to the same category or backgrounds share similar visual patterns. The proposed top-K selection may therefore retain incorrect cross-image connections, while FBS may incorrectly classify background tokens as foreground. The paper provides insufficient analysis of these failure cases, despite claiming that multi-subject consistency emerges naturally without specialized processing. More systematic evaluation of similar-looking subjects, overlapping objects, and complex backgrounds is necessary.

  4. Questionable effectiveness of the adaptive selection mechanism. NSVA adjusts K using a local finite difference of between-class attention variance, but the paper does not establish why maximizing this statistic should improve identity preservation or suppress background leakage. Moreover, the criterion is not guaranteed to identify semantically correct connections, and its effectiveness relative to a carefully optimized fixed-K strategy remains unclear. The proposed temporal caching mechanism also relies on the assumption that selected token indices remain stable across denoising steps, without sufficient quantitative analysis of cache stability, correspondence errors, or their effects on generation quality. Stronger comparisons with fixed-K selection and alternative adaptive strategies are needed.

  5. Evaluation does not fully support the central claims. The main quantitative metrics, CLIP-T, CLIP-I, and DreamSim, measure prompt alignment and perceptual similarity but do not directly quantify the central problem of background leakage or identity confusion. In particular, the claimed ability to preserve multiple distinct subjects is supported mainly by qualitative examples, without dedicated multi-subject identity metrics. Furthermore, the strongest reported results for SelectDiff+ use a different DiT backbone, making it difficult to attribute improvements solely to the proposed method. The paper should provide more comprehensive evaluations of identity confusion, foreground-background consistency, pose diversity, and computational efficiency under matched backbone settings.

Questions for authors

See weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer TMb9

W1. Limited methodological novelty

We agree that top-K selection and the Otsu between-class variance criterion are established tools. Our contribution is a selective feature-interaction mechanism tailored to cross-image subject consistency: selecting reference tokens at the connection level, restricting cross-image interaction at the query level, and adapting the selection budget during denoising. We have distinguished these mechanism designs from the underlying mathematical operations in the contribution statement.

Specifically, SGSA determines which reference tokens each query reads; FBS reuses the selected-score statistics to determine which target queries participate in cross-image interaction; and NSVA uses a local finite difference to update K for the next step. These components control the connection set, the participating queries, and the evolving selection budget, respectively. Reusing similarity statistics avoids a separate mask-extraction procedure, but selection and reduction operations still incur costs; we have therefore removed the phrase “zero additional computation.”

The component ablations in the original Table 2 and the selection-policy comparison in Table R2 below provide empirical support for this design under the evaluated conditions.

W2. Missing recent and closely related baselines

Thank you for identifying these relevant methods. We have expanded the comparison to include CoDi, CharaConsist, ConsiStyle, and SDeC, together with additional evaluation metrics.

Table R1. Expanded comparison with recent methods.

MethodBackboneCLIP-T ↑CLIP-I ↑DreamSim ↓TIFA (%) ↑Aesthetic Score (V2) ↑VLM Score (GPT-6) ↑Time (s) ↓Memory (GB) ↓
PhotoMakerSDXL86.584.739.983.26.371.018.023.7
IP-AdapterSDXL84.694.314.684.65.873.413.519.4
Naive CSASDXL72.574.233.881.44.352.738.042.2
ConsiStorySDXL87.787.431.991.07.280.834.634.5
StoryDiffusionSDXL88.787.532.183.56.478.325.745.6
1Prompt1StorySDXL89.491.219.980.57.166.423.218.7
DreamStorySDXL85.292.516.876.37.273.239.132.5
StoryBoothSDXL88.389.421.482.76.969.227.625.8
MS-DiffusionSDXL84.790.125.886.36.474.514.219.9
CoDiSDXL86.488.121.387.66.981.018.422.5
ConsiStyleSDXL84.288.223.684.26.680.326.833.5
SDeCSDXL87.189.820.986.76.882.520.625.8
CharaConsistDiT90.591.822.588.96.280.629.744.9
SelectDiff (Ours)SDXL90.292.718.591.77.285.219.522.6
SelectDiff+ (Ours)DiT92.194.713.292.17.588.324.630.2

Among the SDXL methods in Table R1, SelectDiff achieves the highest CLIP-T, TIFA, and VLM scores, and ties for the highest aesthetic score. Compared with CoDi, ConsiStyle, and SDeC, it also achieves higher CLIP-I and lower DreamSim. SelectDiff+ obtains better reported quality, runtime, and memory results than CharaConsist. These results strengthen the comparison with recent methods; the matched-checkpoint experiment in Table R4 separately examines the effect of cross-image interaction without changing the backbone.

AuthorsOpenReview ↗

Response to Reviewer TMb9 (Part 2)

W3. Insufficient justification of similarity-based subject selection

We agree that high feature similarity does not necessarily imply correct instance-level identity correspondence. Similarity serves as a signal for selecting cross-image connections, not as a guarantee of identity correctness. Compared with unrestricted feature aggregation, query-specific top-K selection limits the external information received by each query, reducing opportunities for irrelevant or weakly related features to contribute. FBS further restricts which target queries participate in cross-image interaction. These designs aim to reduce incorrect sharing rather than eliminate it in every case.

Incorrect connections can still receive high scores when same-category subjects look similar or backgrounds contain subject-like patterns. Occlusion and styles such as watercolor can further weaken instance discriminability. Supplementary §1.1 discusses these limitations, and Fig. 4 includes a stylized multi-subject failure case. Thus, “without specialized multi-subject processing” describes the implementation, not guaranteed reliability in every multi-subject scenario.

The qualitative comparisons in Fig. 4 show improved separation of subject appearances in the illustrated cases, while the output-level scores in Table R4 provide complementary evidence. Neither constitutes a verification of every selected token correspondence or establishes robustness to all look-alike subjects, occlusions, and complex backgrounds. We have revised the generalization claims accordingly and explicitly distinguished the absence of an additional multi-subject module from guaranteed instance-level matching.

W4. Effectiveness of adaptive selection and temporal caching

Adaptive selection. Between-class variance measures statistical separation between high- and low-similarity scores rather than instance identity. Maximizing it therefore does not guarantee the elimination of leakage or identity confusion. NSVA aims to adapt the selection budget with low search overhead, not to recover an exact semantic boundary.

For a given set of attention scores, all valid K values can be evaluated to obtain the exact single-step variance maximum. Because the scores change during denoising, this is not a global optimum for the generation trajectory or final image quality; NSVA provides no such guarantee either. After sorting, prefix sums allow all candidate objectives to be evaluated in O(L) additional work per score vector. Once the necessary statistics are available, NSVA evaluates the objectives at K and K+1 and applies the update at the next step. This comparison concerns objective evaluation, excluding similarity computation and acquisition of the top-K statistics.

Table R2. Comparison of K-selection policies, using the same cache settings.

K-selection policyCLIP-T ↑CLIP-I ↑DreamSim ↓Time (s) ↓
Validation-selected fixed K85.387.622.417.2
Exact Otsu90.192.319.636.6
NSVA90.292.718.519.5

Relative to the validation-selected fixed K, NSVA increases CLIP-T and CLIP-I by 4.9 and 5.1, reduces DreamSim by 3.9, and adds 2.3 seconds of runtime. Compared with exact Otsu, all three quality metrics are numerically better, while runtime decreases from 36.6 to 19.5 seconds, a reduction of approximately 46.7%. This supports the quality–runtime trade-off of NSVA under the evaluated configuration. It does not imply more accurate optimization of the Otsu objective: exact maximization of a single-step surrogate need not yield better final generation quality.

Temporal caching. We cache selected token indices, while attention scores and features are recomputed from the current state. A stable K alone does not establish stable index sets. We therefore compare recomputing the selection at every step with the current cache policy to assess the end-to-end quality and efficiency effects.

Table R3. Cache ablation with NSVA, Q, and the remaining settings held fixed.

Cache policyCLIP-T ↑CLIP-I ↑DreamSim ↓Time (s) ↓Memory (GB) ↓
No cache; selection recomputed every step90.292.918.447.122.3
Current cache policy90.292.718.519.522.6

Caching reduces runtime from 47.1 to 19.5 seconds, corresponding to approximately 2.42× acceleration. CLIP-T is unchanged, CLIP-I decreases by 0.2, DreamSim increases by 0.1, and memory increases by 0.3 GB. These measurements show a substantial runtime benefit with small differences in the reported quality scores on the evaluated data. They assess the practical trade-off of caching rather than directly proving the stability or semantic correctness of every cached connection.

AuthorsOpenReview ↗

Response to Reviewer TMb9 (Part 3)

W5. Evaluation of the central claims and backbone confounding

We agree that CLIP-T, CLIP-I, and DreamSim do not directly distinguish identity confusion from background leakage. Table R1 adds TIFA, LAION Aesthetic Predictor V2, and GPT-6 composite scores to broaden the assessment of prompt compliance, aesthetic quality, and visual-semantic performance.

We additionally use GPT-6 to assess identity separation and background independence in the generated images. Higher identity-separation scores indicate less confusion between distinct subjects; higher background-independence scores indicate less cross-image background leakage. These are output-level assessments, complementary to global image-similarity metrics.

Table R4. Mechanism comparison using the same FLUX.2 checkpoint.

MethodCLIP-T ↑CLIP-I ↑DreamSim ↓Identity separation ↑Background independence ↑
Independent generation86.742.717.292.396.1
Naive cross-image sharing88.496.127.812.510.9
SelectDiff92.194.713.286.295.7

Independent generation has a CLIP-I score of 42.7, lower than either sharing method, indicating weaker cross-image consistency according to this metric. Its high identity-separation and background-independence scores are not contradictory: avoiding within-image identity confusion and background transfer does not ensure that identities remain consistent across images.

Naive cross-image sharing achieves the highest CLIP-I of 96.1 but has identity-separation and background-independence scores of only 12.5 and 10.9. This illustrates that high global similarity can coexist with unwanted feature sharing. SelectDiff retains a high CLIP-I of 94.7 while increasing the two targeted scores to 86.2 and 95.7. Its DreamSim of 13.2 is also lower than those of independent generation and naive sharing. Although its identity-separation and background-independence scores remain below independent generation, these results support a better balance between cross-image consistency, instance separation, and background independence.

Reviewer_WRHbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Text-to-image diffusion models struggle to guarantee subject appearance consistency across scenes, and unconstrained cross-image attention sharing often leads to background feature leakage. To address this, authors propose SelectDiff, a training-free framework that employs similarity-guided selective attention, retaining only the top-K most similar cross-image token connections to suppress background contamination. To overcome the limitations of a fixed K, authors design a Noise-adaptive Selective Value Adjustment mechanism that dynamically adapts K during the denoising process, and exploit the discrepancy between foreground and background attention distributions. The method outperforms state-of-the-art approaches on the ConsiStory+ benchmark.

Critical strengths

1、The research addresses a valuable problem and the proposed method efficiently resolves the interference of background information on subject features. 2、The overall manuscript exhibits a well-organized structure and coherent logical flow.

Critical weaknesses

1、Comparisons against several SOTA methods such as DreamStory, StoryBooth and MS-Diffusion are absent. These approaches appear capable of alleviating background leakage for subject-consistent generation under the training-free paradigm as well. 2、The qualitative visual results presented in the paper do not seem to match the expected performance claimed by the proposed method and only conventional evaluation metrics are adopted in this work. VQA, aesthetic metrics or VLM-based metrics are not incorporated to comprehensively assess the method.

Questions for authors

Among the favorable cases demonstrated in the paper, all subjects occupy similar fractional areas within the image frame. The authors should present model performance under relatively extreme conditions, such as scenarios where the subject occupies an excessively large or tiny portion of the canvas. It remains unclear whether the selection of Top-K and Top-Q parameters needs to be adapted for such extreme cases.

2 comments
AuthorsOpenReview ↗

Response to Reviewer WRHb

W1–W2. Missing baselines and broader evaluation

Thank you for these suggestions. We have added comparisons with DreamStory, StoryBooth, and MS-Diffusion on the same benchmark, reporting both generation quality and computational performance. To broaden the evaluation beyond the original metrics, we also report TIFA, aesthetic scores, and VLM-based scores for all compared methods. Aesthetic quality is evaluated using LAION Aesthetic Predictor V2, with its raw prediction scores reported. GPT-6 provides a composite assessment of text–image semantic alignment, visual quality, subject consistency, and background leakage. Higher composite scores indicate better overall performance across these dimensions; they should not be interpreted as separate measurements of each dimension.

Table R1. Expanded comparison and additional evaluation metrics.

MethodBackboneCLIP-T ↑CLIP-I ↑DreamSim ↓TIFA ↑Aesthetic Score (V2) ↑VLM Score (GPT-6) ↑Time (s) ↓Memory (GB) ↓
PhotoMakerSDXL86.584.739.983.26.371.018.023.7
IP-AdapterSDXL84.694.314.684.65.873.413.519.4
Naive CSASDXL72.574.233.881.44.352.738.042.2
ConsiStorySDXL87.787.431.991.07.280.834.634.5
StoryDiffusionSDXL88.787.532.183.56.478.325.745.6
1Prompt1StorySDXL89.491.219.980.57.166.423.218.7
DreamStorySDXL85.292.516.876.37.273.239.132.5
StoryBoothSDXL88.389.421.482.76.969.227.625.8
MS-DiffusionSDXL84.790.125.886.36.474.514.219.9
SelectDiff (Ours)SDXL90.292.718.591.77.285.219.522.6
SelectDiff+ (Ours)DiT92.194.713.292.17.588.324.630.2

Among the SDXL methods in Table R1, SelectDiff achieves the highest CLIP-T, TIFA, and VLM scores and ties for the highest aesthetic score. These results support its balance of prompt alignment, visual quality, and subject consistency. Some baselines obtain better scores on individual identity-similarity or efficiency metrics, so we do not claim superiority on every metric. The DiT results provide an additional system-level comparison rather than isolating the effect of the attention mechanism across different backbones.

In our tested multi-subject examples, the newly included baselines also exhibit noticeable identity confusion, similar to the failure patterns illustrated in Fig. 4. We have included these quantitative and qualitative comparisons in the revised manuscript and limited the conclusions to the evaluated cases.

Q1. Very small or large subjects and the choice of K and Q

Thank you for highlighting this condition. K and Q control different aspects of cross-image interaction: K determines the number of reference tokens selected for each query, while Q determines the number of target-image queries that participate in cross-image sharing. NSVA adjusts K during denoising, whereas our Q-rate is 0.8. This permits 80% of target queries to participate in cross-image interaction; the remaining queries retain within-image attention. It does not imply a 20% reduction in total computation.

We stratified the generated images by subject area and report the results below.

Table R2. Performance stratified by the fraction of image area occupied by the subject.

Subject sizeSubject area fractionImagesCLIP-T ↑CLIP-I ↑DreamSim ↓
Small<20%67290.2792.7118.44
Medium20%–80%153990.3892.7718.42
Large>80%97390.3192.7418.47

The reported scores are close across the three groups: the maximum differences are 0.11 for CLIP-T, 0.06 for CLIP-I, and 0.05 for DreamSim. This provides evidence of consistent performance across the evaluated area groups. It does not establish that NSVA finds an optimal K, or that all extremely small-subject cases are covered by the small-subject group. We have included this analysis and clarified the scope of the robustness claim in the revised manuscript.

Dear Reviewer,

Thank you again for your constructive comments and valuable evaluation of our work.

We would like to kindly invite you to take another look at our responses and reconsider your assessment in light of our clarifications. We would greatly appreciate it if you could let us know whether our responses have adequately addressed your concerns, or if there are any remaining questions that we could further clarify.

Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Area Chairs and All Reviewers

We thank the reviewers for their thoughtful feedback and recognition of SelectDiff’s strengths. We appreciate JE2v’s positive assessment of the method’s simplicity and elegance, the insight of using query–key similarity as an implicit subject mask, and the creative perspective of variance-based adaptive K. WRHb highlights the practical value of addressing background interference and the paper’s clear organization, while TMb9 recognizes the computational appeal of selective sharing and its balance of identity preservation, prompt alignment, and efficiency. Building on this feedback, we have clarified the technical contributions and implementation details and added comparisons and analyses to address the reviewers’ concerns. These changes are highlighted in red in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.