acceptodds
Under review as a conference paper at ICLR 2027

Harmonia: Optimizing Aesthetic Gains within the Semantic Elasticity Boundary for Aesthetic Photo Reconstruction

Abstract

Aesthetic Photo Reconstruction (APR) improves photo aesthetics through structural edits while preserving the original subjects and scene. However, APR inherently faces an aesthetic–semantic tension: larger structural changes may yield greater aesthetic gains but also increase the risk of altering subject identity and scene semantics. We therefore formulate APR as maximizing aesthetic gains within a photo-dependent semantic elasticity boundary. Existing methods optimize reconstruction plans only in text space, without feedback from their realized visual outcomes, making it difficult to assess semantic feasibility and optimize aesthetics within this boundary. They also lack support for intent-conditioned personalized APR. To address these limitations, we propose Harmonia, a two-stage APR framework: (1) Structured Photographic Planning Learning, which decomposes reconstruction plans into atomic photographic operations and jointly learns reference-free, reference-guided, and intent-conditioned planning; and (2) Feasible-Space Visual Optimization, which introduces Visual Aesthetic GRPO (VA-GRPO) to execute candidate plans and optimize the planner from their realized visual outcomes within the semantic elasticity boundary. We further construct Harmonia-Bench for evaluating both default and personalized APR. Experiments show that Harmonia outperforms AesFormer using only 30% of its training data even without fine-tuning the image editor, while additionally supporting personalized APR for the first time.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aesthetic Photo Reconstruction, Photographic Aesthetics, Multimodal Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_8qDAOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The author proposed a two-stage training pipeline (SFT + GRPO) to train an image editing planner to produce image editing instructions to improve the aesthetics of the image without large changes in the semantics.

Critical strengths
  • The user study helps strengthen the claim and avoids circularity of using the same rubric for training and eval.
  • The support for personalized APR is a useful new domain, and it respects the individual taste differences and could help avoid homogenizing the output.
  • The semantic gate effectively limits the changes to the image and keeps the SFR high.
Critical weaknesses

One of the paper’s main claims, that text-based reward cannot capture feasibility, is never tested. It is possible the semantic gate (identity, clothing, physical scene, and key content) could be measured from the text space alone (eg if the text space says “replace the cloth with a red T-shirt”). The authors do not need to rerun the GRPO using text space, but a simple check could help strengthen the claim: use the generated pairs of images; use the image semantic gate as GT to test if an LLM can detect the same pattern in the text space with a prompt similar to “Could this editing instruction change the semantics of the image in [list of gates]”. Similarly, the aesthetics gain might also be able to infered from the text domain. I can think of two ways text-based reward could fail, and authors could include them in the motivation if the text-based reward has low agreement: 1. the image editing model is not perfect; it might change semantics when asking for non-semantic changes (for example, it might change the cloth every time you ask it to change the angle), using GRPO could teach the LLM to avoid these failure areas. 2. Some prompts might not seem to change semantically but would in practice, and using an image to verify is easier and more accurate (for example, asking the model to change the angle might seem like a non-semantic change, but it would because the model needs to outpaint some areas) This is my only concern for the paper and I am willing to raise my score if this is solved. It is fine if the evidence cannot support that image-level reward is necessary, as long as the authors tighten the claim.

Questions for authors

See weaknesses. Could the author provide a small test comparing image and text space reward?

4 comments
AuthorsOpenReview ↗

(1/2) Response to Reviewer 8qDA

Thank you for your valuable comments. Following your suggestion, we conducted two experiments to test whether predictions from the reconstruction plan (without the edited image) can replace (1) semantic feasibility checks and (2) visual aesthetic reward evaluation.

  • Setup: We use Qwen3.7-Plus (Harmonia's Gate and reward judge) and GPT-5.6 (an independent, strong model). Both receive only the original image and Harmonia's text plan; predictions are compared with scores obtained from the actual edited images (GT).
  • Data: Samples come from 55,296 real editing rollouts collected during VA-GRPO training, each recording the planner's instruction, FLUX.2's output image, and the resulting evaluation scores. Thus, we directly test whether pre-edit predictions can reproduce the semantic Gate decisions and visual rewards during VA-GRPO training.

1. Text-Based Predictions Cannot Replace the Visual Semantic Gate

  • Experiment: To test both (1) violation detection and (2) preservation of feasible edits, we sample 2,000 infeasible reconstructions (500 per violation category: Identity, Clothing, Physical Scene, Key Content) and 2,000 matched feasible reconstructions. From the original image and plan, each model predicts the same four 1–5 scores as the visual Gate, using the image-based Gate decisions as GT.

  • Metrics:

    • Balanced Accuracy averages violation and feasible recall (50% random baseline);
    • Violation Miss Rate is the fraction of violations incorrectly accepted;
    • Feasible Recall is the fraction of feasible edits correctly accepted;
    • MCC measures prediction–GT association (0: none; 1: perfect).
    ModelBalanced Acc. ↑Violation Miss ↓Feasible Recall ↑MCC ↑
    Qwen3.7-Plus57.10%72.05%86.25%0.175
    GPT-5.655.48%15.90%26.85%0.134
  • Conclusion: Balanced Accuracy is only 55.48%–57.10%, near chance (50%). Qwen3.7-Plus incorrectly accepts 72.05% of actual violations; GPT-5.6 detects more violations but incorrectly rejects 73.15% of feasible edits. On additional 4,799 naturally distributed rollouts sampled to preserve the original feasibility distribution, Balanced Accuracy remains only 58.64%/57.68%. Identity/Clothing violation miss rates on this set are 98.18%/94.08% (Qwen3.7-Plus) and 87.27%/81.58% (GPT-5.6), indicating that both struggle to detect violations. Thus, text-based predictions cannot reliably detect semantic violations while retaining feasible edits; they cannot replace the visual Gate.

2. Text-Based Predictions Cannot Reliably Estimate Visual Rewards

  • Experiment: We sample 600 complete 8-rollout groups (4,800 edits, all passing the visual Gate), stratified by early/middle/late training steps and Default/Personalized tasks. Using only the original image and plan, both models predict R1 (Structural Gain), R2 (Final Aesthetics), R3 (Intent Satisfaction), and their combined reward, against image-based GT. We also test 600 naturally distributed 8-rollout groups containing feasible and infeasible edits: predict the Gate first, then rewards, to reconstruct VA-GRPO's final reward and ranking.

  • Metrics: Pairwise Accuracy measures correct ordering of candidate pairs within each 8-rollout group (50% random baseline); Spearman measures predicted–GT rank correlation (0: none; 1: perfect). We also report Top-1 Accuracy (selecting the best of eight) and MAE (average absolute score error) in the discussion below.

    TargetQwen Pairwise ↑GPT Pairwise ↑Qwen Spearman ↑GPT Spearman ↑
    R1: Structural Gain51.26%51.91%0.1090.257
    R2: Final Aesthetics48.84%49.58%-0.069-0.139
    R3: Intent Satisfaction59.71%59.13%0.3800.361
    Combined Reward52.80%52.77%0.2360.287
  • Conclusion: R1/R2 pairwise accuracy is only 48.84%–51.91% (50% random). For R2, Spearman is negative for both models, showing that predicted scores fail to track actual aesthetic quality; MAEs are 12.03/19.84 points (out of 100). Selecting the best-looking edit among eight succeeds only 11.83%/12.17% of the time (12.50% random). Although R3 retains some intent signal (~59% pairwise), combined rewards reach just 52.80%/52.77%. On the naturally distributed groups, jointly predicting Gate and rewards yields only 52.60%/50.52% final-reward pairwise accuracy (50% random) and 15.00%/13.50% Top-1 accuracy (12.94% random). Thus, text-based predictions cannot reliably rank actual edits or provide the aesthetic reward signal needed by VA-GRPO.

These results confirm Harmonia's motivation: a text plan specifies intended edits, not their realized semantic changes or aesthetic quality. VA-GRPO therefore evaluates actual edited images for both feasibility and reward optimization.

AuthorsOpenReview ↗

(2/2) Response to Reviewer 8qDA

3. Why Text-Based Rewards Fail

  • Analysis: Building on the reviewer's insightful observations, we further illustrate two reasons why text-based rewards can fail:

    • (1) Limited editor capabilities: Current image editors may struggle to understand or execute complex photographic instructions. For example, FLUX.2 often fails to follow "Set the bottom edge of the frame just below the woman's feet," but responds better to "Reframe as a full-body portrait, with both feet fully visible." Similarly, "Shoot from a high angle" often produces an excessively steep viewpoint, while "Slightly raise the camera viewpoint" better achieves the intended adjustment. Such failures depend on the editor's capabilities and limitations, and cannot be reliably predicted from textual instructions alone.
    • (2) Unintended semantic changes: Even seemingly reasonable instructions, such as changing the camera viewpoint, may require generating previously unseen background regions based on the editor's learned scene priors. For example, when changing the viewpoint of a portrait taken outside a café, the editor may introduce tables and chairs based on typical café layouts, even if they were absent from the original scene. These outcomes are difficult to predict from textual plans alone.
  • Solution: APR requires the planner not only to determine how to improve photographic aesthetics, but also to ensure that the editor can execute these changes accurately and appropriately. Accordingly, Harmonia's planner provides professional aesthetic reasoning, while VA-GRPO uses actual visual feedback to improve aesthetic planning and adapt to the editor's capabilities and limitations. This enables the planner to learn executable reconstruction plans that achieve aesthetic gains while preserving semantic consistency. We will incorporate this analysis into the final paper.

My concerns are resolved, I will raise my score

AuthorsOpenReview ↗

Response to Reviewer 8qDA's Follow-up Comment

Thank you for your constructive feedback and positive response. We are glad that our additional experiments and analysis have addressed your concerns. We will incorporate these results and the corresponding analysis into the final version of the paper.

Reviewer_1ZM7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Harmonia, a VLM-based framework for aesthetic photo reconstruction by changing photographic structures, such as viewpoint, composition, and human pose, while keeping subject identity and scene semantics largely unchanged. The framework first decomposes reconstruction plans into 6 atomic operations, then executes candidate plans and uses GRPO to optimize the planner based on evaluation across multiple dimensions. The authors also introduce Harmonia-Bench, covering synthetic and real photos for general and personalized reconstruction tasks. Experiments show strong gains over existing methods on aesthetic photo reconstruction.

Critical strengths
  1. The motivation for evaluating the actual visual reconstruction instead of text instruction is meaningful, as text space is less informative of the aesthetic and semantic validity of the actual visual outcome. In addition, treating semantic validity as a constraint rather than a reward is reasonable, which encourages exploring more diverse candidate reconstructions within a semantically valid range.

  2. The experimental evaluation is quite comprehensive, showing improvement over various baselines in both automatic and human evaluations.

Critical weaknesses

The semantic elasticity boundary is not explicitly modeled. As a central contribution that motivates the paper, the boundary is implemented as a binary gate with a fixed threshold across four dimensions. The method does not estimate a photo-dependent boundary or characterize the permissible degree of modification. It remains unclear how this formulation differs fundamentally from conventional semantic preservation constraints in image editing.

Questions for authors
  1. How sensitive is VA-GRPO to the semantic gate threshold (currently 4/5)? An ablation with different thresholds would help show the robustness of the design choice.

  2. In Table 1, the improvements in Aesthetic Quality (S2) over the strongest baselines are relatively small, particularly on AesRecon. Are the improvements statistically significant? Could it be due to the different image editor, and how do the authors interpret the gap between Structural Improvement (S1) and Aesthetic Quality (S2)?

  3. The six atomic operations appear to overlap (e.g. Camera, Composition, and Framing), and Preservation is more of an editing constraint than an operation. Could the authors clarify the distinction between these dimensions and show whether each contributes meaningfully to planning performance?

4 comments
AuthorsOpenReview ↗

(1/2) Response to Reviewer 1ZM7

Thank you for your valuable comments. Below we respond to the comments in Weaknesses (W) and Questions (Q).

1. Semantic Elasticity Boundary (W1)

  • Why SEB is not explicitly modeled: The semantically feasible space in APR is high-dimensional, irregular, and discontinuous, and the feasibility of a reconstruction plan depends on its actual visual outcome. Explicitly modeling the entire boundary would require (1) exploring nearly infinite plans and (2) executing them to obtain their actual outcomes, which is computationally expensive and unnecessary. Crucially, determining whether a reconstruction lies within the feasible space does not require knowing its complete boundary in advance. Accordingly, Harmonia uses a semantic gate to determine whether each candidate reconstruction IeiI_e^i lies within the feasible space of its original photo IpI_p: Gi=I[min⁡d∈Ssdsem(Ip,Iei)≥4]G_i=\mathbb{I}\left[\min_{d\in\mathcal{S}}s_d^{\mathrm{sem}}(I_p,I_e^i)\ge4\right], enabling efficient, photo-dependent constrained optimization without explicitly modeling the entire boundary.

  • Difference from conventional image editing: The key distinction lies in the optimization objective. (1) Conventional image editing modifies specified content while preserving unedited content as much as possible. (2) In contrast, APR requires substantial global changes to achieve significant aesthetic improvements, making it inappropriate to maximize semantic consistency. Accordingly, Harmonia treats semantic preservation as a feasibility constraint rather than a reward to maximize, enabling more extensive structural adjustments. Table 2 (main paper) shows that, compared with using a Consistency Reward, Harmonia improves S1, S2, and S3 by an average of 0.73, 1.00, and 2.26, respectively, validating this design.

2. Semantic Gate Threshold Sensitivity (Q1)

Experiment 1: Human Agreement

  • Experiment: We selected over 600 original–reconstruction pairs, covering low, moderate, and high semantic consistency (approximately 200 pairs each). Five annotators independently evaluated each pair, with majority voting determining the human judgment. We compared the agreement between different Gate thresholds(3/4/5) and human judgments:

    Gate ThresholdAgreement with Human Judgments
    ≥372%
    ≥4 (Ours)95%
    ≥583%
  • Conclusion: Threshold 3 is too permissive (72% agreement), allowing semantically inconsistent reconstructions to pass; threshold 5 is too restrictive (83% agreement), rejecting many acceptable reconstructions. We therefore adopt threshold 4, which achieves the highest agreement with human judgments (95% agreement).

Experiment 2: Gate Threshold Sensitivity

  • Experiment: Starting from the same Stage 1 model, we independently retrained Stage 2 using Gate thresholds of 3, 4, and 5, and evaluated Structural Improvement (S1), Aesthetic Quality (S2), and Semantic Feasibility Rate (SFR) on PARA Default:

    Gate ThresholdS1 ↑S2 ↑SFR ↑
    ≥377.2358.1264.89%
    ≥4 (Ours)75.1857.8991.19%
    ≥571.0854.6292.86%
  • Conclusion: Compared with threshold 4, threshold 3 improves S1/S2 by 2.05/0.23 but reduces SFR by 26.30 pp, compromising semantic preservation. Threshold 5 improves SFR by only 1.67 pp while reducing S1/S2 by 4.10/3.27, limiting aesthetic improvements. Threshold 4 achieves a better balance between aesthetic gains (S1/S2: 75.18/57.89) and semantic preservation (SFR: 91.19%).

AuthorsOpenReview ↗

(2/2) Response to Reviewer 1ZM7

3. Statistical Significance of S2 (Q2)

  • Experiment: We conducted two-sided paired t-tests on per-image S2 scores against AesFormer, with Holm correction for multiple comparisons. Wilcoxon signed-rank tests yielded consistent results.

    SettingΔS2 ↑95% CIAdjusted p-value
    AesRecon Default+0.88[+0.35, +1.41]0.002
    AesRecon Personalized+0.05[-0.29, +0.39]0.768
    PARA Default+3.82[+3.21, +4.43]<0.001
    PARA Personalized+3.51[+3.05, +3.98]<0.001
  • Conclusion: Harmonia achieves statistically significant S2 improvements in 3 of the 4 settings (AesRecon Default, PARA Default, and PARA Personalized). We further analyze below.

  • Analysis: S1 measures photographic structural improvement, whereas S2 measures overall aesthetic quality, which is also influenced by color and lighting. AesFormer explicitly includes Color & Light in its planning, optimizing both structure and appearance. In contrast, Harmonia targets realistic reshooting, focusing on composition, viewpoint, and pose while intentionally preserving the original color and lighting style. Consequently, Harmonia achieves larger S1 improvements, while its S2 gains are relatively smaller. Nevertheless, Harmonia achieves the highest S2 across all four settings, using only approximately 30% of AesFormer's training image pairs even without fine-tuning the image editor.

4. Different Image Editors (Q2)

  • Experiment: To control for editor differences, we executed reconstruction plans generated by Harmonia and AesFormer using both FLUX.2 (Harmonia's editor) and AesEditor (AesFormer's editor), and evaluated the results under identical settings on PARA:

    EditorPlannerDefault S1 ↑Default S2 ↑Personalized S1 ↑Personalized S2 ↑Personalized S3 ↑
    FLUX.2AesFormer68.7755.8067.4654.7264.75
    FLUX.2Harmonia (Ours)75.1857.8974.0357.6174.51
    AesEditorAesFormer61.7054.0765.5954.1054.10
    AesEditorHarmonia (Ours)75.5856.3674.1255.9264.82
  • Conclusion: Harmonia outperforms AesFormer on all 5 metrics with either editor. Across both editors, Harmonia improves S1 by 6.41–13.88, S2 by 1.82–2.89, and S3 by 9.76–10.72. These results demonstrate that Harmonia's advantage stems from stronger photographic planning rather than differences in image editors.

5. Distinction and Effectiveness of the Six Atomic Operations (Q3)

  • Distinction: The six operations serve distinct functions: (1) Camera controls shooting distance and viewpoint; (2) Pose controls subject posture and actions; (3) Composition controls subject placement and spatial relationships; (4) Framing controls image boundaries and subject visibility (e.g., setting the bottom frame boundary at mid-thigh); (5) Exclusion specifies distracting elements to avoid; and (6) Preservation specifies essential content to retain. The first four govern photographic structure, while the last two govern content selection and preservation. Table 2 (main paper) shows that, compared with removing the atomic-operation structure, the full Stage 1 improves S1, S2, and S3 by an average of 10.11, 1.03, and 11.52, respectively, validating the effectiveness of structured planning.

  • Effectiveness:

    • Experiment: We conducted per-dimension ablations on 200 randomly sampled reconstruction plans generated by Harmonia on PARA. For each ablation, we manually removed instructions corresponding to one dimension while keeping all others unchanged, then executed the modified plans using the same FLUX.2 editor:

      ConfigurationS1 ↑ΔS1S2 ↑ΔS2
      Harmonia (Ours)76.34—57.86—
      w/o Camera73.38-2.9656.21-1.65
      w/o Pose72.00-4.3453.21-4.65
      w/o Composition74.81-1.5356.35-1.51
      w/o Framing74.63-1.7156.29-1.57
      w/o Exclusion73.84-2.5056.31-1.55
      w/o Preservation74.78-1.5656.35-1.51
    • Conclusion: Removing any dimension degrades reconstruction quality, reducing S1 by 1.53–4.34 and S2 by 1.51–4.65. Removing Pose has the largest impact, decreasing S1/S2 by 4.34/4.65. These results demonstrate that all six operations contribute meaningfully to reconstruction quality.

I'd like to thank the authors for their response and the additional results. Most of my concerns have been addressed, and I will keep my positive rating.

AuthorsOpenReview ↗

Response to Reviewer 1ZM7's Follow-up Comment

Thank you for your positive feedback. We are glad that our clarifications have addressed your concerns. We sincerely appreciate your constructive comments and continued support.

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

Harmonia addresses Aesthetic Photo Reconstruction, where a system should improve photographic composition and appearance without changing the subject or scene. Its first stage trains a planner to express edits as six types of atomic photographic operations under reference-free, reference-guided, and intent-conditioned settings. The second stage, VA-GRPO, samples eight plans per training example, executes them with a frozen image editor, rejects results that fail a four-part semantic-consistency gate, and rewards structural improvement, final aesthetics, and intent satisfaction. On Harmonia-Bench, Harmonia performs strongly on both the AesRecon-derived synthetic subset and the PARA portrait subset.

Critical strengths
  • The most compelling design choice is to optimize the planner using the edited images rather than judging proposed edits only in text. This directly addresses the gap between a plausible plan and its actual visual outcome.
  • Decomposing a plan into atomic operations makes the output easier to inspect and execute. The same framework also handles default and intent-conditioned reconstruction.
  • Treating semantic preservation as a feasibility requirement is well motivated. Figure 3 shows that removing the gate increases the aesthetic scores but sharply reduces the semantic feasibility rate.
Critical weaknesses
  • The paper describes the semantic elasticity boundary as photo-dependent, but its operational definition always uses the same four criteria and requires every score to be at least 4. The scores are conditioned on each image pair, but the criteria and cutoff do not adapt across images. A threshold-sensitivity study would clarify how much the result depends on this fixed choice.
  • AesFormer retains its original editor while the other methods use the same FLUX.2 editor. Table 1 is therefore a practical system comparison rather than a controlled comparison of planning quality, which weakens the claim of outperforming AesFormer.
  • Planning, semantic gating, structural-gain scoring, and editing rely on several external or proprietary models. The paper does not show whether the ranking is stable when the gate, reward judge, aesthetic model, or editor is changed.
  • VA-GRPO executes eight edited candidates for every training example, and the reported experiments use ten A40 GPUs, yet wall-clock time and image-generation cost are not reported.
Questions for authors

Please refer to the Weaknesses section.

3 comments
AuthorsOpenReview ↗

(1/2) Response to Reviewer fpEu

Thank you for your valuable comments. Below we respond to the comments in Weaknesses (W).

1. Unified Evaluation Criteria (W1)

  • Analysis: We respectfully clarify that the photo-dependence of SEB does not require different evaluation criteria across photos. The core requirement of APR is consistent: preserving the original (1) subject and (2) scene (see main paper, line 66). Accordingly, the Semantic Gate evaluates (1) subject consistency (Identity, Clothing) and (2) scene consistency (Physical Scene, Key Content), thereby covering both core semantic preservation requirements of APR. For each candidate reconstruction IeiI_e^i, the Gate evaluates its consistency with the original photo IpI_p: Gi=I[min⁡d∈Ssdsem(Ip,Iei)≥4]G_i=\mathbb{I}\left[\min_{d\in\mathcal{S}}s_d^{\mathrm{sem}}(I_p,I_e^i)\ge4\right]. Since the semantic scores depend on the specific content of each photo, the same structural adjustment may be feasible for one photo but violate the semantics of another. Thus, the unified criteria cover APR's subject-and-scene preservation requirements while accommodating photo-dependent feasible spaces.

2. Threshold-sensitivity Study (W1)

Experiment 1: Human Agreement

  • Experiment: We selected over 600 original–reconstruction pairs, covering low, moderate, and high semantic consistency (approximately 200 pairs each). Five annotators independently evaluated each pair, with majority voting determining the human judgment. We compared the agreement between different Gate thresholds(3/4/5) and human judgments:

    Gate ThresholdAgreement with Human Judgments
    ≥372%
    ≥4 (Ours)95%
    ≥583%
  • Conclusion: Threshold 3 is too permissive (72% agreement), allowing semantically inconsistent reconstructions to pass; threshold 5 is too restrictive (83% agreement), rejecting many acceptable reconstructions. We therefore adopt threshold 4, which achieves the highest agreement with human judgments (95% agreement).

Experiment 2: Gate Threshold Sensitivity

  • Experiment: Starting from the same Stage 1 model, we independently retrained Stage 2 using Gate thresholds of 3, 4, and 5, and evaluated Structural Improvement (S1), Aesthetic Quality (S2), and Semantic Feasibility Rate (SFR) on PARA Default:

    Gate ThresholdS1 ↑S2 ↑SFR ↑
    ≥377.2358.1264.89%
    ≥4 (Ours)75.1857.8991.19%
    ≥571.0854.6292.86%
  • Conclusion: Compared with threshold 4, threshold 3 improves S1/S2 by 2.05/0.23 but reduces SFR by 26.30 pp, compromising semantic preservation. Threshold 5 improves SFR by only 1.67 pp while reducing S1/S2 by 4.10/3.27, limiting aesthetic improvements. Threshold 4 achieves a better balance between aesthetic gains (S1/S2: 75.18/57.89) and semantic preservation (SFR: 91.19%).

3. Controlled Comparison with AesFormer (W2)

  • Experiment: To isolate the effect of photographic planning, we executed reconstruction plans generated by Harmonia and AesFormer using both FLUX.2 (Harmonia's editor) and AesEditor (AesFormer's editor), and evaluated the results under identical settings on PARA:

    EditorPlannerDefault S1 ↑Default S2 ↑Personalized S1 ↑Personalized S2 ↑Personalized S3 ↑
    FLUX.2AesFormer68.7755.8067.4654.7264.75
    FLUX.2Harmonia (Ours)75.1857.8974.0357.6174.51
    AesEditorAesFormer61.7054.0765.5954.1054.10
    AesEditorHarmonia (Ours)75.5856.3674.1255.9264.82
  • Conclusion: Harmonia outperforms AesFormer on all 5 metrics with either editor. Across both editors, Harmonia improves S1 by 6.41–13.88, S2 by 1.82–2.89, and S3 by 9.76–10.72. These results demonstrate that Harmonia's advantage stems from stronger photographic planning rather than differences in image editors.

AuthorsOpenReview ↗

(2/2) Response to Reviewer fpEu

4. Robustness to External Models (W3)

  • Experiment: To evaluate robustness to external models, we replaced both the Semantic Gate and reward judges with Gemini 3 Flash, and independently retrained Stage 2 from the same Stage 1 model. We evaluated all four settings on AesRecon and PARA:

    AesReconDefault S1 ↑Default S2 ↑Personalized S1 ↑Personalized S2 ↑Personalized S3 ↑
    Baseline (Qwen3-VL-8B)55.2338.3057.0737.6458.34
    Harmonia (Ours)76.5848.6475.2947.0763.21
    Harmonia (Replaced Gate & Reward Judges)77.3747.8576.1546.6361.05
    PARADefault S1 ↑Default S2 ↑Personalized S1 ↑Personalized S2 ↑Personalized S3 ↑
    Baseline (Qwen3-VL-8B)55.2152.3355.5552.0553.41
    Harmonia (Ours)75.1857.8974.0357.6174.51
    Harmonia (Replaced Gate & Reward Judges)75.4657.8874.7457.1668.07
  • Conclusion: After replacing both the Gate and reward judges, Harmonia maintains comparable improvements. Across the 10 evaluation metrics, 4 improve over the original Harmonia, and all 10 outperform the baseline. These results demonstrate that VA-GRPO remains effective across different external models. Meanwhile, the choice of judges still affects performance: our original configuration achieves higher S3 on both AesRecon (63.21 vs. 61.05) and PARA (74.51 vs. 68.07).

  • Additional Evaluation: We respectfully note that a complete Stage 2 training run requires approximately 92 hours on 10 A40 GPUs. Given this computational cost and the two additional threshold-sensitivity training runs, we prioritized full retraining with alternative Gate and reward judges. Editor robustness is separately validated in W2 above. We will include further cross-model evaluations in the final paper.

5. Computational Cost (W4)

  • A complete Stage 2 (VA-GRPO) training run takes approximately 92 hours (3.8 days) on 10 NVIDIA A40 GPUs (48 GB each). The wall-clock time is broken down as follows:

    StageTimeProportion
    Image Generation42 h45.7%
    GRPO Optimization45 h48.9%
    Reward Evaluation5 h5.4%
    Total92 h100%
  • Conclusion: Image generation and GRPO optimization account for the majority of computational costs, contributing 45.7% and 48.9% of the total training time, respectively.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer fpEu,

Thank you once again for your valuable comments. As the discussion phase is approaching its end, we would like to kindly ask whether our clarifications and additional experiments have sufficiently addressed your concerns. We would be happy to provide further clarification if needed.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.