acceptodds
Under review as a conference paper at ICLR 2027

NextShot: Learning Source-Grounded Structured Composition Targets for Portrait Photography

Abstract

Modern visual models increasingly move beyond recognizing isolated content toward structured understanding of scene entities, semantics, and relationships. Relational information is useful not only for characterizing what is observed: when a task requires inferring a desired configuration from existing visual content, that configuration often depends on interactions among entities and their global context. We study this problem in photographic composition, where relevant scene elements are already present and composition depends on how people and environmental regions are positioned and scaled relative to one another and the frame. We formulate Region-Level Composition Target Prediction, which preserves source-region identities while jointly predicting regional geometry, target frame aspect ratio, and person framing. Because photographic demonstrations typically provide only whole-image source–reference outcomes, direct region-level supervision is unavailable. We recover structured supervision by transferring reference geometry to source identities through partial correspondence while preserving each reference as a complete alternative arrangement, complemented by single-image aesthetic annotations. Our predictor, NextShot, models region–region and region–frame relations, integrates multimodal scene semantics, and conditions regional geometry on predicted frame and person framing. NextShot improves centroid and visible-area prediction over an autoregressive alternative. Two independent large multimodal evaluators from different model families also consistently prefer its complete layouts over alternative predictors and key model variants. In a 100-case guided-recapture study, five independent human evaluators prefer photographs taken with NextShot guidance in 78.0% of 500 judgments. These results establish source-grounded structured targets as a learnable representation for explicit spatial goal prediction from relational scene information.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Photographic Composition, Structured Prediction, Relational Learning, Image Aesthetics, Photography Guidance

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NtBgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Region-Level Composition Target Prediction, a task that predicts the desired spatial arrangement of semantic regions in a photograph. The predicted targets include region bounding boxes, centroids, visible areas, frame aspect ratio, and person framing.

To construct training supervision, the authors derive AesRecon-CT from existing source-reference photograph pairs. They establish partial region correspondences between source and reference images and transfer the reference geometry to the corresponding source regions.

The proposed model, NextShot, combines three components: a region-frame relational encoder, graph-conditioned multimodal fusion with a pretrained VLM, and a framing-conditioned geometry decoder. The model is jointly trained using reference-derived geometric supervision and auxiliary aesthetic annotations from PortraitCraft.

Experiments demonstrate improvements in geometric prediction accuracy over several baselines. Two VLM-based evaluators generally prefer the predicted layouts from NextShot over alternative methods. Furthermore, a guided photography study reports that human evaluators prefer photographs taken with NextShot guidance in 78% of judgments.

Critical strengths
  1. The task formulation is interesting and potentially useful to real-world applications.
  2. The source-reference supervision construction is reasonably designed.
  3. The experimental evaluation is relatively comprehensive.
Critical weaknesses
  1. User study contains only five people, which might be biased.
  2. The improvements in geometric accuracy are relatively modest, and there is apparently some trade-offs.
  3. Since the task is to give a better camera pose, why not just use bbox? This paper argues that these attributes capture complementary geometric information, but there seems no ablation tests on it, such as bbox-only, centroid-and-area, and the full representation. Ideally, if bbox prediction is accurate, it should be more than enough to guide photographer. More importantly, these variants should ideally be evaluated through composition preference rather than geometric accuracy alone, since their output spaces are different.
  4. The VLM-based layout evaluation lacks human validation. The paper relies heavily on two VLM evaluators to assess the quality of predicted layouts and demonstrate the effectiveness of the architectural components. While the paper conducts a human preference study on the actual photographs before and after guided recapture, this does not validate the VLM-based evaluation of the predicted layouts. In particular, no human evaluation is provided to verify whether the improvements over architectural ablations are consistent with human preferences. This is concerning because the two VLM evaluators sometimes disagree substantially, such as in the comparisons with NoVLM-AttrAux and NextShot-8B. I would like to see a human preference evaluation on a subset of the predicted layouts and an analysis of human–VLM agreement.
Questions for authors

See weakness 3 and 4.

5 comments
AuthorsOpenReview ↗

Response to Reviewer NtBg (1/3)

Thank you for recognizing the merits of our work and for your careful review and constructive feedback. We address your concerns in turn below.

W1. Size of the user study

Thank you for pointing out the limited size of the user study. We used several complementary evaluations to examine different aspects of NextShot’s effectiveness. We added a human evaluation of predicted layouts to assess whether the model’s composition targets reflect people’s preferences and are meaningful in practice. The recapture study also helped us assess how well the human judgments carry over to actual photographs and whether photographers can use NextShot’s spatial guidance to improve their shots.

We agree that a systematic study with a more diverse group of photographers, subjects, and evaluators, across a wider range of shooting conditions, would help establish how broadly the method applies.

W2. Modest gains in geometric accuracy and trade-offs across metrics

Thank you for highlighting the trade-offs among the geometric metrics. We believe these arise in part from the nature of the composition task. The metrics measure different properties of a predicted layout, including each region’s spatial extent, the centroid of its visible content, its visible area, and the aspect ratio of the frame. Improvements in these properties need not occur together.

Geometric metrics primarily measure how closely a prediction matches a reference composition. They do not directly measure the quality of the composition as a whole. We therefore assess NextShot using regional agreement with the reference geometry, preferences for the overall layout, and the quality of the resulting photographs. This gives a broader assessment of composition than any single geometric error. In direct comparisons of the predicted layouts, human evaluators showed a clear preference for the full model.

AuthorsOpenReview ↗

Response to Reviewer NtBg (2/3)

W3. Why predict bounding boxes, visible centroids, and visible areas jointly? Are bounding boxes alone sufficient?

Thank you for pointing out that the choice of geometric representation needed further validation. We agree. We trained a BBox-only model and used geometric metrics, model evaluations, and human layout comparisons to test whether bounding boxes alone suffice to predict the composition.

(1) Are bounding boxes alone sufficient?

We ran additional ablations to address this question directly.

ModelOverall box IoU ↑Person box IoU ↑Environment box IoU ↑Centroid MAE ↓Area MAE ↓Frame aspect-ratio error ↓
Full NextShot0.63730.59730.66300.06590.06704.07%
BBox-only0.63620.59980.6600——3.79%
Centroid-and-area———0.06650.06343.31%

To examine what these geometric fields contribute to the overall composition, we asked Qwen, Flash, and human evaluators to compare the predicted layouts.

EvaluatorFull NextShotBBox-onlyTie
Qwen58.00%35.75%6.25%
Flash53.50%33.00%13.50%
Human evaluators221 (73.7%)7 (2.3%)72 (24.0%)

We also examined the layouts predicted by the two models in individual scenes. We found that NextShot better captured spatial relationships between objects. For example, in a scene with a person and a chair, BBox-only assigned the chair a relatively large region, making the chair appear too large relative to the person. NextShot predicted a more plausible relative scale and placement for the person and the chair, producing a more coherent layout.

The value of this design goes beyond adding prediction fields or inputs. By jointly modeling the spatial extents of regions and incorporating information such as scene depth, NextShot can better capture and predict how people and their surroundings are arranged.

AuthorsOpenReview ↗

Response to Reviewer NtBg (3/3)

W4. Do human judgments support the VLM layout evaluations? Do people and models show the same overall preferences?

Thank you for raising this question. Other reviewers also raised this issue and suggested repeating the evaluation with candidate positions swapped. We revised the Flash evaluation to use the same position assignments as Qwen. To test the stability of the results, we then reversed the candidate order within each pair and evaluated the same pairs again.

(1) Do the main model comparisons remain stable under the revised protocol?

For the same set of test pairs, we first placed NextShot in position A for 100 pairs and in position B for the other 100. We then swapped the two candidates within each pair and repeated the evaluation.

Compared methodFlash win rateFlash win rate after swap
NoChange59.5%66.5%
AR-8B72.5%76.0%
NoVLM-AttrAux53.5%52.5%
w/o AesAux73.5%71.0%
w/o RelBias67.5%66.5%
w/o Framing Conditioning54.5%51.5%

For NoVLM-AttrAux, balancing candidate positions substantially narrowed the gap between the two evaluators, although some disagreement remained. We examine this comparison further below using the human layout judgments.

(2) How does presentation order affect model preferences?

Our observations are consistent with earlier work on position bias in multi-image vision-language models. Tian et al. [1] found that the position of an image can influence these models’ judgments. Hu et al. [2] evaluated each candidate pair in both orders to reduce the influence of presentation order on multimodal preference judgments. These studies help explain the sensitivity we observed and support evaluating preferences in both orders. Although the models’ layout preference rates varied, the overall pattern continued to favor our method.

(3) Do human evaluators also prefer the full model’s layouts?

To test directly whether the predicted layouts reflect human preferences for composition, we invited three professional photographers to assess the same 100 test scenes in a blind evaluation. They considered the person’s position and size, the relationship between the person and the surroundings, and the overall arrangement within the frame.

Method compared with full NextShotPreferred full modelPreferred other modelRated equally
NoVLM-AttrAux164 (54.7%)24 (8.0%)112 (37.3%)
w/o Framing Conditioning208 (69.3%)13 (4.3%)79 (26.3%)
NextShot-8B202 (67.3%)11 (3.7%)87 (29.0%)
BBox-only221 (73.7%)7 (2.3%)72 (24.0%)

The photographers’ preferences provide further support for the full model and its core design choices. The full model’s advantage over NoVLM-AttrAux was limited in the automated evaluation. This suggests that NoVLM-AttrAux already produces strong structured layouts that differ only slightly from those of the full model. Even so, the photographers cast more votes for the full NextShot model when they compared the layouts directly.

We also evaluated both presentation orders to examine why Qwen and Flash originally favored different models in the 4B versus 8B comparison. Flash selected the 4B model in 56% of judgments and the 8B model in 36%. The human evaluators also showed a clear preference for the 4B model in the layout comparison. Both the automated judgments and the photographers’ preferences support our choice of the 4B configuration. For this task and training setup, a larger backbone offered no clear gain in layout preference.

These additional experiments give us a fuller picture of the automated evaluation results and further support the full NextShot model’s advantage in the main architectural comparisons. We will include these findings in the revised manuscript.

Thank you again for your detailed and constructive feedback. We hope these revisions address your concerns and give you greater confidence in the technical soundness and research contributions of our work.

References:

[1] Tian et al., “Identifying and Mitigating Position Bias of Multi-image Vision-Language Models,” CVPR 2025.

[2] Hu et al., “Multimodal RewardBench 2: Evaluating Omni Reward Models for Interleaved Text and Image,” CVPR 2026.

Reviewer_PRzEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NextShot, which formulates portrait photography guidance as a region-level structured target prediction task, jointly predicting the positions and scales of person and environmental regions, as well as the target frame aspect ratio. The authors construct training supervision through partial region correspondence between source and reference images, and combine relational modeling, multimodal features, and auxiliary aesthetic supervision for prediction.

Critical strengths

The paper formulates photographic composition optimization as a region-level spatial target prediction task, explicitly specifying how people and environmental elements should be repositioned and resized. Compared with conventional textual guidance or image cropping, this provides more concrete and actionable composition recommendations.

The authors transfer geometric information from high-quality reference photographs to source-image regions through partial region correspondence, while preserving the integrity of alternative reference compositions. This provides a feasible approach to learning multiple valid photographic layouts.

NextShot outperforms NoChange and AR-8B on the applicable geometric prediction metrics. The real-world photography study further demonstrates the practical potential of the method, with guided recaptures receiving 78.0% of human preference judgments across 100 photography cases.

Critical weaknesses
  1. The reliability of automatic composition evaluation requires further validation.

The paper primarily relies on Qwen and Flash to evaluate the composition quality of predicted layouts, but the two evaluators show substantial disagreement in several key comparisons. For example, in Table 1, the win rates of NextShot against NoVLM-AttrAux are 45.0% and 72.5%, respectively. Similarly, the comparison between the 4B and 8B models in Appendix B.2 reveals opposite preference trends.

Furthermore, Appendix F.1 indicates that Flash always places NextShot in position A during test-set layout evaluation, without swapping candidate positions for the same samples. This may introduce positional bias. Since automatic layout preferences constitute an important part of the evaluation, additional position-swapping experiments and human layout evaluations would help establish the stability and consistency of the results.

  1. The real-world photography study lacks sufficient controls to clearly attribute the observed improvements to NextShot.

In Section 5.4, photographers first capture an initial photograph independently and then take another photograph after viewing the composition targets predicted by NextShot. Although the guided photographs receive 78.0% of human preference judgments, the experimental design does not rule out improvements resulting from repeated shooting or photographers' own adjustments.

Moreover, the study does not include unguided recaptures or comparisons with alternative composition guidance methods, making it difficult to isolate the benefits of NextShot's specific guidance over general composition suggestions. The experiment also involves only four photographers and two subjects, limiting the diversity of subjects and shooting conditions. Further evaluation is needed to establish the generalizability of the method.

  1. The quality of the automatically constructed region-level supervision is insufficiently validated.

The paper constructs training supervision using pretrained detection and segmentation models together with heuristic region matching. However, it does not provide quantitative analyses of region correspondence accuracy, matching coverage, or major error categories. In particular, large environmental regions such as the sky and ground may exhibit ambiguous segmentation boundaries and semantic correspondences, potentially affecting the accuracy of the target geometry.

Additionally, the authors propose preserving complete reference compositions to avoid inconsistencies caused by mixing different layouts, but do not directly compare this strategy with reference averaging or per-region target selection. Human verification of a sampled set of region correspondences and additional ablation experiments would help establish the reliability and effectiveness of the structured supervision.

Questions for authors

Do the main conclusions remain consistent when the candidate positions are swapped during evaluation?

Could the authors include unguided recaptures or comparisons with other composition guidance methods to further validate the practical contribution of NextShot?

Have the region correspondences between source and reference images been manually verified? Could the authors report matching accuracy and coverage statistics?

7 comments
AuthorsOpenReview ↗

Response to Reviewer PRzE (1/3)

Thank you for recognizing the merits of our work and for your careful review and feedback. We address each question below.

W1 / Q1. Do layout preferences hold after a position swap? How can we explain the disagreement between evaluators on NoVLM-AttrAux and on the 4B versus 8B comparison?

Following your suggestion, we revised the protocol for evaluating both presentation orders. On the same set of pairs, we first placed NextShot in position A for 100 pairs and in position B for the other 100. We then reversed the presentation order within each pair and repeated the evaluation.

(1) Do the main layout preferences hold after swapping positions?

Compared methodNextShot win rate (Flash)Flash tie rateNextShot win rate after swap (Flash)Flash tie rate after swap
NoChange59.5%0.5%66.5%1.0%
AR-8B72.5%2.5%76.0%2.0%
NoVLM-AttrAux53.5%7.5%52.5%12.0%
w/o AesAux73.5%7.5%71.0%5.5%
w/o RelBias67.5%14.5%66.5%10.0%
w/o Framing Conditioning54.5%17.0%51.5%16.5%

After the swap, evaluators generally still favored NextShot over the main baselines and core variants. As a further check, Qwen evaluated the same pairs in both orders and also favored NextShot over AR-8B, w/o AesAux, and w/o RelBias.

Some win rates varied with presentation order, consistent with recent work on position bias. Tian et al. [1] found that image position can affect judgments by multi-image vision-language models. Hu et al. [2] evaluated each candidate pair in both orders to reduce the influence of order on multimodal preference judgments.

(2) Do people also prefer the full model’s layouts?

To test human preferences, three professional photographers assessed the layouts in a blind evaluation. They considered the person’s position and size, the relationship between the person and the surroundings, and the overall arrangement within the frame. The results are shown below.

Method compared with NextShot-4BPreferred full modelPreferred other methodRated equally
NoVLM-AttrAux164 (54.7%)24 (8.0%)112 (37.3%)
w/o Framing Conditioning208 (69.3%)13 (4.3%)79 (26.3%)
NextShot-8B202 (67.3%)11 (3.7%)87 (29.0%)

The human judgments further support the full model’s layout advantage. We found subtle differences between layouts from NoVLM-AttrAux and the full model. The automated evaluators did not always agree on these subtle differences. The full model arranged the person and surroundings more coherently in terms of their relative scale and placement. For example, in some scenes both methods predicted similar visible areas for the person but allocated different amounts of space to the surrounding regions. The human preference for the full model further supports jointly modeling region relationships when predicting the overall layout.

(3) Do the additional 4B and 8B results support our model choice?

Qwen and Flash originally favored different models in the 4B versus 8B comparison. We therefore swapped candidate positions for Flash on each of the original 100 development images. With both rounds weighted equally, Flash chose 4B in 56% of judgments and 8B in 36%, with 8% ties. Presentation order affected the original evaluation. Across both orders, Flash favored 4B. The human evaluators also clearly preferred 4B in the layout comparison. A larger backbone offered no clear gain in layout preference for this task and training setup.

Thank you again for suggesting these experiments. They helped us examine our main conclusions more thoroughly.

References:

[1] Tian et al., “Identifying and Mitigating Position Bias of Multi-image Vision-Language Models,” CVPR 2025.

[2] Hu et al., “Multimodal RewardBench 2: Evaluating Omni Reward Models for Interleaved Text and Image,” CVPR 2026.

AuthorsOpenReview ↗

Response to Reviewer PRzE (2/3)

W2 / Q2. Does NextShot guidance offer a benefit beyond an unguided recapture? Do the current results show that it outperforms other composition guidance methods?

Following your suggestion, we added 50 real-world scenes and included an unguided recapture in each as a direct control to assess the practical value of the guidance.

(1) Does NextShot guidance still improve composition beyond what photographers achieve with an unguided recapture?

Five independent evaluators compared the photographs from the three conditions in a blind pairwise evaluation. Flash and Qwen also evaluated the composition of the same photographs.

Within-scene comparison (first vs. second)Human judges preferred firstHuman judges preferred secondRated equallyFlash preferred firstQwen preferred first
Unguided recapture vs. initial photograph168 (67.2%)48 (19.2%)34 (13.6%)80%74%
Guided recapture vs. unguided recapture171 (68.4%)42 (16.8%)37 (14.8%)72%84%
Guided recapture vs. initial photograph207 (82.8%)31 (12.4%)12 (4.8%)92%97%

The most important comparison is between the guided and unguided recaptures. Human evaluators already preferred the unguided recaptures to the initial photographs, showing that photographers can improve composition by adjusting the camera position and framing on their own. When the two recaptures were compared directly, evaluators more often preferred the NextShot-guided photographs. Flash and Qwen showed the same preference, selecting the guided recapture in 72% and 84% of judgments, respectively.

(2) Is this preference also evident across different scenes?

To see whether a few scenes accounted for the overall vote, we also examined the majority preference within each scene. In the direct comparison, a majority of evaluators preferred the NextShot-guided photograph in about 40 of the 50 scenes. The human evaluators and both model evaluators showed the same overall preference. We believe this provides further evidence of NextShot’s value.

(3) Number of participants, other forms of guidance, and scope of the study

We agree that the numbers of photographers, subjects, and scenes remain limited. The study therefore gives only a partial picture of performance across different populations and under complex shooting conditions. We expanded participation and the range of shooting settings as far as we could. We also assessed the method through human preferences for the actual photographs, human comparisons of predicted layouts, and evaluations by two models. The new unguided recapture control in particular allowed us to test more directly what NextShot adds to photographers’ own adjustments. An important next step is to conduct controlled studies with more diverse photographers and subjects across a wider range of scenes and shooting conditions.

AuthorsOpenReview ↗

Response to Reviewer PRzE (3/3)

W3 / Q3. Reliability of region-level supervision and effectiveness of keeping each reference composition intact

Following your suggestion, we manually checked region matches and measured matching coverage. We also compared our strategy with two alternatives, averaging targets across references and selecting targets per region.

(1) How reliable are the region correspondences, and what proportion of the source regions do they cover?

We sampled 100 source–reference image pairs from the training set. Three evaluators each checked every pair. Their assessments indicate that our strategy matches regions with high precision.

Verification metricResult
Matching precision95.92%
95% confidence interval[94.48%, 97.25%]
Recall of matchable regions76.91%
95% confidence interval[72.88%, 80.79%]

We also measured matching coverage by region type in the training set.

Region typeMatched regionsTotal source regionsMatching coverage
Person regions4057474285.55%
Environmental regions83301169771.21%
All regions123871643975.35%
Sky (separate analysis)1258135193.12%

Although the source and reference photographs in each pair come from the same session, the photographer may change the camera position, shooting direction, or framing to emphasize the subject and reduce distractions. This changes how much of the background and individual regions are visible. When constructing the dataset, we manually filtered out pairs whose backgrounds differed too much for reliable matching. We applied a conservative partial matching strategy to these regions. When a correspondence was uncertain or ambiguous, we left it unmatched and transferred no geometric supervision.

During manual verification, we found that small objects such as ornaments and shoulder bags were the most difficult to match. Occlusion and changes in pose or viewpoint can alter their visible extent or segmentation. The results for sky, which the reviewer specifically mentioned, support the method’s ability to identify and match large environmental regions.

(2) Why keep each reference intact? How does this strategy compare with averaging across references or selecting a target per region?

We keep each reference intact because different photographs can show several valid compositions of the same scene. In each photograph, the person’s position and scale, the share of the frame occupied by the environment, and the frame aspect ratio together form an observed layout. We examined this rationale in the 101 development scenes with multiple references. Even allowing for ties, in 62 scenes (61.4%) no reference minimized geometric error for all evaluated regions at once. A local optimum for each region therefore need not give the best overall layout.

We then conducted ablation experiments.

Geometric metricComplete referenceCross-reference averagingPer-region reference selection
Overall box IoU ↑0.637340.633560.63618
Person box IoU ↑0.597280.592970.59493
Environment box IoU ↑0.663050.659900.66178
Visible-centroid MAE ↓0.065940.065920.06600
Visible-area MAE ↓0.067010.068020.06699
Frame aspect-ratio relative error ↓4.07%4.30%3.64%

To examine the effect on the overall composition, we asked Qwen and Flash to compare the layouts predicted by the corresponding models.

Comparison strategyEvaluatorComplete reference preferredComparison strategy preferredTie rate
Cross-reference averagingQwen43.25%45.00%11.75%
Flash43.50%26.00%30.50%
Per-region reference selectionQwen52.00%34.75%13.25%
Flash33.50%29.00%37.50%

In the comparison with per-region reference selection, both evaluators more often chose layouts from the model trained on complete references. This agrees with our observations on scenes with multiple references.

The manual checks support the reliability of region correspondence supervision, and the direct ablations against the two alternatives further support keeping each reference intact. Improving the reliability of correspondences under complex viewpoint changes and partial occlusion remains a direction for future work.

Thank you for your helpful feedback. We hope these additional experiments and revisions address your concerns and further demonstrate the value of our work. In light of these results, we hope you will consider raising your overall rating.

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

This paper formulates region-level composition target prediction. Given a portrait and its source regions, the model jointly predicts identity-preserving boxes, visible centroids, visible-area fractions, frame aspect ratio, and person framing, so that the output serves as an explicit spatial goal for the next shot rather than a critique of the current one.

The supervision is built by transferring reference geometry onto source-region identities through partial correspondence, with each reference kept intact as one complete alternative. Attribute scores and rationales from PortraitCraft are used only as auxiliary signal. Architecturally, the model combines region–frame relational encoding, multimodal fusion, and decoding that conditions regional geometry on predicted frame and person framing. Evaluation covers reference agreement, layout preference, and real guided recapture.

Critical strengths

S1. The formulation is the real contribution. What I find most useful is that predictions stay tied to the observed regions. That makes the output inspectable and directly usable by a photographer, and it keeps regional geometry and global framing coupled instead of treating them as separate regressions.

S2. The supervision design fits the task. Matching on appearance and semantics rather than position and scale is appropriate for this task, since composition change is exactly what alters position and scale. Keeping each reference whole also avoids producing averaged layouts that no photograph actually demonstrated.

S3. Structured decoding works and is fast. On the fields both methods predict, it beats AR-8B on centroid and visible-area error, with paired intervals supporting the difference. The reported median service time is 0.206 s, compared with 16.522 s for AR-8B. However, these timings use prepared inputs and exclude upstream perception.

S4. The recapture study provides preliminary evidence of practical usability. A 78% preference across 500 judgments, with majorities in 89 of 100 pairs, suggests photographers can act on these targets. I would not overread it, but it does show the representation is usable outside benchmark metrics.

Critical weaknesses

W1. The aesthetic claims rest too heavily on model judges. Against lightweight predictors and ablations, the evidence is mostly preference rates from Qwen and Flash. The NoVLM-AttrAux comparison is the clearest problem: The reported preference rate is 45.0% under Qwen and 72.5% under Flash. Flash also sees NextShot fixed in position A, and no human layout comparison is reported. I would want position-swapped judgments, and ideally human comparisons, before accepting that the full architecture produces better compositions.

W2. The recapture gain is not cleanly attributable. The study compares the initial shot with a second shot taken after the photographer receives guidance, so part of the improvement may simply come from shooting again. An unguided retake, or retakes under alternative guidance, would separate these effects. With four photographers and two subjects, I also treat the result as a demonstration rather than evidence of broad generality.

Questions for authors
  1. Flash always sees NextShot as candidate A in the test-layout comparison, and each pair is judged only once, so I cannot tell how much of the reported preference is positional. Would it be possible to rerun the comparison with the two candidates swapped and report both the resulting rates and the proportion of pairs whose judgment flips?

  2. The NoVLM-AttrAux comparison is the one result I would most like clarified: 45.0% under Qwen versus 72.5% under Flash is a large gap. Is it concentrated in one of the four criteria, or spread across all of them?

  3. Geometric metrics for the full model and the lightweight predictors are close, so the argument for the full architecture rests almost entirely on model-judge preferences. Given that the paper already runs a human study for photographs, would a small human comparison on predicted layouts be practical?

5 comments
AuthorsOpenReview ↗

Response to Reviewer osxC (1/3)

Thank you for recognizing the merits of our work and for your careful review and thoughtful feedback. We address each of your questions below.

W1 / Q1 / Q3. Are the evaluation results stable when candidate positions are swapped? Does the human evaluation support the full model’s advantage in layout prediction?

Thank you for this suggestion. We revised how we evaluate the two presentation orders. In the first round, we placed NextShot in position A for 100 pairs and in position B for the other 100. We then swapped the candidates within each pair and repeated the evaluation so that every pair was assessed in both orders.

(1) Evaluation results after swapping candidate positions

Compared methodWin rate before swapTie rate before swapWin rate after swapTie rate after swapInconsistency rate
NoChange59.5%0.5%66.5%1.0%30.0%
AR-8B72.5%2.5%76.0%2.0%13.0%
NoVLM-AttrAux53.5%7.5%52.5%12.0%36.5%
w/o AesAux73.5%7.5%71.0%5.5%25.5%
w/o RelBias67.5%14.5%66.5%10.0%22.0%
w/o Framing Conditioning54.5%17.0%51.5%16.5%21.0%

In some comparisons, the win rate varied with the order in which the candidates were presented. This is consistent with recent work on position bias in multi-image vision-language models. Tian et al. [1] found that an image’s position in a sequence can influence these models’ judgments. In their multimodal preference evaluation, Hu et al. [2] assessed each candidate pair in both orders to reduce the influence of presentation order. In our additional experiments, the overall direction of preference remained the same in the main model comparisons, even though some win rates varied. In Q2, we discuss where the two evaluators disagreed on NoVLM-AttrAux and what may explain the difference.

As a further check, Qwen evaluated the same pairs in both orders. Its judgments also support NextShot’s advantage over the main baselines and core model variants.

(2) Human evaluation of predicted layouts

To assess how well the predicted layouts reflect human preferences for composition, we invited three professional photographers to evaluate the layouts in a blind comparison. To judge which layout had the better composition, they considered the person’s position and size, the relationship between the person and the surroundings, and the overall arrangement within the frame.

Compared methodPreferred full NextShotPreferred the other methodRated equally
NoVLM-AttrAux164 (54.7%)24 (8.0%)112 (37.3%)
w/o Framing Conditioning208 (69.3%)13 (4.3%)79 (26.3%)
NextShot-8B202 (67.3%)11 (3.7%)87 (29.0%)

The human evaluation further supports the full model’s advantage in layout prediction. Together, these findings support NextShot’s ability to predict the overall layout and the core design choices in its architecture.

Thank you again for your suggestions.

References:

[1] Tian et al., “Identifying and Mitigating Position Bias of Multi-image Vision-Language Models,” CVPR 2025.

[2] Hu et al., “Multimodal RewardBench 2: Evaluating Omni Reward Models for Interleaved Text and Image,” CVPR 2026.

AuthorsOpenReview ↗

Response to Reviewer osxC (2/3)

Q2. Where do the evaluators disagree in the comparison with NoVLM-AttrAux?

Thank you for raising this question. As reported in W1, we balanced the candidate positions and evaluated each pair in both orders. Flash selected NextShot in 53.5% of comparisons before the swap and 52.5% after it, for an average of 53.0%.

To understand the small remaining gap between the evaluators, we examined the predicted layouts more closely. NoVLM-AttrAux already produces structured layouts with plausible compositions, so the overall differences from the full model are relatively small. The evaluators do not always judge these subtle differences consistently. In the full model’s layouts, however, the relative scale and placement of the person and the surrounding elements are more harmonious. For example, in some scenes the two methods predicted similar visible areas for the person but allocated different amounts of space to the surrounding regions. We believe the additional human evaluation helps us interpret the disagreement more accurately and assess what the full model adds beyond existing structured prediction methods.

AuthorsOpenReview ↗

Response to Reviewer osxC (3/3)

W2. Does NextShot guidance offer a compositional benefit beyond an unguided recapture?

Thank you for this suggestion. We added 50 real-world scenes and included an unguided recapture as a direct control.

(1) Does NextShot guidance still improve composition beyond what photographers achieve with an unguided recapture?

In the additional experiment, the same photographer took both recaptures of the same subject in each scene. For the unguided recapture, the photographer used their own judgment to adjust the camera position and framing. For the guided recapture, the photographer made adjustments using the region boxes and semantic information provided by NextShot. Five independent evaluators made blind pairwise comparisons of the photographs from the three conditions. We also used Flash and Qwen to evaluate the composition of the same photographs.

Within-scene comparisonHuman judges preferred firstHuman judges preferred secondRated equallyFlash preferred firstQwen preferred first
Unguided recapture vs. initial photograph168 (67.2%)48 (19.2%)34 (13.6%)80%74%
Guided recapture vs. unguided recapture171 (68.4%)42 (16.8%)37 (14.8%)72%84%
Guided recapture vs. initial photograph207 (82.8%)31 (12.4%)12 (4.8%)92%97%

The most important comparison is the one between guided and unguided recaptures. Human evaluators more often preferred the unguided recaptures to the initial photographs, indicating that photographers can improve composition by adjusting the camera position and framing on their own. When the two recaptures were compared directly, evaluators more often preferred the NextShot-guided photographs.

(2) Is this preference also evident across different scenes?

A majority of evaluators preferred the NextShot-guided photograph in 39 of the 50 scenes. This supports the overall vote totals and shows that the preference for guided recaptures extended beyond a few scenes.

(3) Number of participants, other forms of guidance, and scope of the study

We agree that the study still includes a limited number of photographers, subjects, and scenes. We added several experiments to assess NextShot’s practical value from different perspectives. An important direction for future work is to include a more diverse group of photographers and subjects and to study a wider range of shooting environments and conditions.

Thank you again for your constructive feedback. We hope the additional experiments and corresponding revisions address your concerns and further demonstrate the value of our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

We sincerely thank the chairs for the time and effort devoted to organizing and coordinating the review process. We also thank the reviewers for their thoughtful feedback and careful review of our paper, responses, and additional experiments. As the review discussion period draws to a close, we provide the following brief summary to help the chairs and reviewers gain an overview of the paper’s main strengths and the progress of the discussion.

We appreciate the reviewers’ recognition of the following main strengths:

  1. Novelty of the task formulation and target representation (osxC, PRzE). For example, Reviewer osxC acknowledged the value of keeping targets tied to source regions and jointly representing regional geometry and global framing, stating, “The formulation is the real contribution.”

  2. Well-motivated supervision design (osxC, PRzE, NtBg). All three reviewers recognized the merits of the supervision design; Reviewer osxC and Reviewer PRzE highlighted partial region correspondences and the preservation of complete reference compositions.

  3. Experimental performance and practical usability (osxC, PRzE, NtBg). Reviewer osxC and Reviewer PRzE acknowledged the prediction performance and preliminary evidence of practical usability provided by the recapture study. Reviewer osxC also recognized prediction efficiency, while Reviewer NtBg noted practical potential and relatively comprehensive evaluation.

In response to the reviewers’ suggestions, we made additions and revisions. We summarize several key updates below:

  1. Evaluation reliability and human validation (osxC, PRzE, NtBg). We balanced candidate positions, evaluated each pair in both orders, added blind layout comparisons by professional photographers, and analyzed disagreements between evaluators.

  2. Attribution of recapture gains and study size (osxC, PRzE; Reviewer NtBg also raised concerns about study size). We added scenes, an unguided recapture control, and independent blind human evaluations to assess NextShot’s contribution beyond photographers’ own adjustments.

  3. Supervision reliability and complete-reference training (PRzE). We added manual checks of region correspondences, coverage and error analyses, and ablations with cross-reference averaging and per-region reference selection.

  4. Joint geometric representation and metric trade-offs (NtBg). We added BBox-only and Centroid-and-area ablations, compared the full model with BBox-only through human and model-based layout evaluations, and clarified the distinction between geometric agreement and overall composition preference.

After reviewing our responses and additional evidence, all three reviewers explicitly stated that their questions or concerns had been addressed, and Reviewer PRzE also raised their rating. We sincerely appreciate the reviewers’ further feedback during the discussion:

ReviewerRatingFollow-up feedback (verbatim)
osxC3 (Weak Accept)“Thank you for your detailed response. My question is now resolved.”
PRzE2 → 3 (Weak Accept)“Thank you for the detailed reply. My question has been resolved.” “Based on the clarifications and additional evidence, I have raised my rating to Weak Accept.”
NtBg3 (Weak Accept)“Thanks for the detailed rebuttal. My concerns have been addressed.”

Once again, we sincerely thank all chairs and reviewers for their time, patience, and support. We have incorporated the additional experiments, analyses, and necessary clarifications into Sections 3.2, 5.3, and 5.4 of the main text and Appendix D, among other parts of the revised manuscript. Please see our revised manuscript for details.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.