acceptodds
Under review as a conference paper at ICLR 2027

Mirrorrefine: Mirror Reflection Generation via Multi-agent Refinement with Spatial Priors

Abstract

Advancing visual generation toward world modeling requires more than photorealism: generated views must remain consistent with the underlying scene. In mirror-reflection synthesis, this requires inferring reflected content and spatial layout from the source scene, beyond the text–image alignment emphasized by general-purpose prompt-refinement methods. We propose MirrorRefine, a multiagent framework for mirror-reflection generation through multi-action evolutionary refinement. Starting from a generated or supplied image, the Reflection Optimization Controller analyzes the original request and visible content outside the mirror to infer intended reflected content and construct action-specific synthesis conditions, rather than treating potentially erroneous reflections as reliable targets. The Quality Evaluator provides a vision–language model with coarse reflectionlayout priors predicted by an offline-trained Reflection Spatial Geometry (RSG) module. Together with textual conditions, these priors guide the diagnosis of geometric, appearance, and illumination inconsistencies in the synthesized image. Based on this diagnosis, the evaluator selects among image generation, mirrorregion inpainting, and targeted editing, and returns structured revision suggestions to the controller. The controller updates the synthesis conditions and coordinates the selected operation, after which the resulting image is reassessed to guide subsequent refinement. All component parameters, including those of RSG, remain fixed during iterative refinement. Experiments on MirrorBenchV2, MSD, and CCMirror show complementary strengths: among the compared systems, the online configuration leads the reported geometry, lighting, and appearance scores, while the local configuration achieves the highest PSNR on MSD and CC-Mirror.

Primary area:
generative models
Keywords:
Training-Free, Image Generation

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_GnxrOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

MirrorRefine is an inference-time multi-agent framework for mirror-reflection generation and completion with frozen generation models. It combines Source–Mirror–Reflection reasoning to infer reflected content with an offline-trained Reflection Spatial Geometry (RSG) module that predicts coarse reflection locations. A VLM-based evaluator diagnoses geometry, appearance, and illumination errors, selects synthesis actions, and verifies candidates within a budget of 8 synthesis attempts. Experiments show complementary contributions from reflection context and RSG. The online configuration leads the reported GLAM-Bench scores, while the local configuration achieves the highest PSNR on MSD and CC-Mirror; both configurations trail PhysReflect in PSNR and IoU on MirrorBenchV2.

Critical strengths
  1. Complementary contributions of reflection reasoning and spatial guidance MirrorRefine uses Source–Mirror–Reflection reasoning to infer reflected content and RSG to predict reflection locations. In Table 3, removing reflection context causes larger drops in appearance and lighting scores than removing RSG, whereas removing RSG causes larger drops in geometry scores, IoU, and BF1. This ablation supports the use of both components to address different aspects of reflection consistency.

  2. Completion quality on real-image datasets With frozen generation models, MirrorRefine-local achieves the highest PSNR among the compared methods on MSD and CC-Mirror, and the best SSIM and LPIPS on CC-Mirror. These gains in reference-based metrics support improved completion quality on real images, independently of the VLM-based evaluation.

Critical weaknesses
  1. Limited evidence of advantages over mirror-specific methods Table 1 highlights scores within groups, placing MirrorRefine alongside general-purpose inference-time refinement methods. This presentation does not emphasize that PhysReflect outperforms both MirrorRefine configurations on PSNR, LPIPS, IoU, and BF1 on MirrorBenchV2. MirrorRefine itself uses reflection-trained RSG and Source–Mirror–Reflection reasoning and is evaluated in single-planar-mirror settings, so comparisons with mirror-specific methods are particularly relevant. Its advantages over these methods in Table 1 lie mainly in GLAM-Bench scores. However, GLAM-Bench uses GPT-5.6 Sol, also the online configuration’s controller and evaluator, which may bias the evaluation in favor of that configuration. Its geometry scores favor MirrorRefine-online over PhysReflect, contrary to the IoU and BF1 rankings. The user study includes only five participants and reports neither agreement among participants nor confidence intervals, limiting the strength of this evidence. Computational costs are also not reported, as discussed in Weakness 2. PhysReflect performs better on reference-image and annotation-based geometry metrics on MirrorBenchV2, while the evidence favoring MirrorRefine relies mainly on a potentially biased VLM evaluation and a small user study without uncertainty estimates. These results do not provide a strong reason to choose MirrorRefine over PhysReflect for this completion task.

  2. Insufficient isolation of the agent loop’s contribution and cost-effectiveness Appendix Table 5 compares MirrorRefine-local with its standalone inpainting tool, FLUX.1-Fill-dev, only on selected geometry metrics, with limited and inconsistent gains. PSNR, LPIPS, and GLAM-Bench comparisons are missing. Table 3 supports the contributions of reflection context and RSG, but all variants retain the iterative loop and use FLUX.1-Fill-dev. Even “Without both” retains action selection and candidate retention. These ablations therefore do not isolate the full loop’s benefit over a single inpainting call. Table 1 also varies backbones, synthesis budgets, and candidate-selection procedures. A controlled comparison using the same generation model and matched computational budgets is needed against simpler alternatives: generating multiple FLUX.1-Fill-dev candidates and selecting one with a VLM, or repeatedly regenerating images with prompts revised from VLM feedback. Neither comparison is provided. The synthesis budget counts only image-generation attempts. VLM call counts, token usage, and execution time for the controller and evaluator are not reported. The reflection-specific agent loop is the paper’s central contribution, but comparisons with the standalone tool are limited, and controlled comparisons with simpler alternatives at the same cost are missing. Therefore it remains unclear whether the added complexity of the proposed framework provides meaningful benefits.

Questions for authors

A. Questions that could materially affect the assessment

  1. (Weakness 2) Could you report the results of a single standalone execution of FLUX.1-Fill-dev on the main metrics in Table 1, including PSNR, SSIM, LPIPS, IoU, BF1, and GLAM-Bench?
  2. (Weakness 2) Using the same generation model (FLUX.1-Fill-dev) and the same total computational budget, could you provide comparisons against the following simpler alternatives? (e.g. Generating multiple candidates and using a VLM to select one / Having a VLM identify corrections without the proposed agent structure, incorporating its feedback into the prompt, and repeating generation)
  3. (Weaknesses 1 and 2) Could you report the number of VLM calls made by the controller and evaluator, token usage, and execution time per sample? Comparisons with mirror-specific methods such as PhysReflect and other inference-time methods would be helpful.
  4. (Weakness 1) Could you provide GLAM-Bench results rescored by a model that is not used in any MirrorRefine configuration?
  5. (Weakness 1) Could you provide evidence supporting the reliability of the user-study results, such as agreement among participants (inter-rater agreement) and confidence intervals? Could you also clarify how participants were recruited, including their relationship to the authors, without compromising review anonymity?

B. Clarification requests

  1. Are the controller’s Analysis Controller, Reflection Content Inference, and Prompt Rewriter implemented as separate VLM calls or combined into a single call? Are they executed again at every round? Could you provide the exact prompts and example outputs for each stage?
  2. What algorithm is used to construct M^{err} and M^{target} in practice?
  3. Which generation models are used for the TIR, T2I-Copilot, RAISE, and Noise Scaling baselines?
  4. How are reflected-object masks extracted from generated images when computing IoU and BF1?
10 comments
AuthorsOpenReview ↗

## W1 / A4 / A5-1: Advantages over mirror-specific methods and evaluation reliability

PhysReflect's stronger reference-based geometry results on MirrorBenchV2 are obtained with mirror-task training on SynMirrorV2, whose held-out scenes form MirrorBenchV2, and rendered ground-truth depth outside the mirror as an additional generation input. In contrast, MirrorRefine performs training-free refinement using frozen pretrained synthesis backbones with a single mirror-erased RGB image and mirror mask, without depth input or mirror-task fine-tuning of the synthesis backbones.

Training and input differences. MirrorRefine's RSG is trained separately on Mirror-Reflect-Inference and remains frozen during refinement. We additionally retrained PhysReflect and MirrorVerse without depth inputs, retaining their architectures and original training settings.

MethodIoU ↑BF1 ↑GLAM-Geo ↑
MirrorVerse without depth0.41640.131560.14
PhysReflect without depth0.46110.139361.05
MirrorRefine-local0.42730.167369.54
MirrorRefine-online0.40690.147774.43

PhysReflect's IoU/BF1 decrease from 0.6932/0.4228 to 0.4611/0.1393 without depth, narrowing its IoU advantage. Both MirrorRefine configurations achieve higher BF1 and GLAM-Geometry than the depth-free baselines, while PhysReflect retains higher IoU. These results suggest that access to additional rendered ground-truth depth contributes substantially to the stronger MirrorBenchV2 geometry performance of depth-conditioned mirror-specific methods. Such accurate depth is difficult to obtain in real-world scenes, where MirrorRefine achieves better reflected-object IoU/BF1 than the evaluated specialized baselines.

Objective geometry on real images. We additionally evaluated reflected-object IoU and BF1@2px on MSD and CC-Mirror using the same evaluation procedure across methods (B4).

MethodMSD IoU ↑MSD BF1@2px ↑CC IoU ↑CC BF1@2px ↑
PhysReflect0.11670.03010.08050.0375
PhysReflect (MSD fine-tuned)0.18270.0477——
RAISE0.14800.04180.13800.0484
MirrorRefine-online0.21510.05440.24410.0771

Online achieves the highest IoU and BF1@2px among these methods on both real-image datasets, including against PhysReflect fine-tuned on MSD. These measurements provide reference-based geometry evidence beyond VLM evaluation.

Independent VLM evaluation (A4). We rescored the same 500 MirrorBenchV2 final outputs using DeepSeek-V4.1-Flash-1017 and GPT-6.1 Sol, neither of which is used in MirrorRefine's configurations.

JudgeMethodGeometry ↑Lighting ↑Appearance ↑
DeepSeek-V4.1-Flash-1017PhysReflect61.8879.6677.88
Qwen-controller hybrid66.8584.0286.35
MirrorRefine-online68.2384.7988.78
GPT-6.1 SolPhysReflect81.0286.3179.02
Qwen-controller hybrid82.3388.1787.54
MirrorRefine-online85.1889.6890.68

Both independent judges rank Online above PhysReflect and the Qwen-controller hybrid across all three dimensions, with Geometry margins of 6.35 and 4.16 points over PhysReflect. This reduces the concern that the reported GLAM advantage is specific to the original GPT-5.6 Sol judge.

For the original real-image GLAM evaluation, the valid/unscored counts per method and dimension are 200/0 on MSD and 199/1 on CC-Mirror. The single unscored CC-Mirror case resulted from a content-policy refusal and was excluded from the common comparison set.

AuthorsOpenReview ↗

## W1 / A4 / A5-2: Advantages over mirror-specific methods and evaluation reliability

Human evaluation reliability (A5). We conducted a separate blinded, geometry-focused pairwise study with 15 participants and 200 image pairs (3,000 valid judgments), comparing MirrorRefine-online against PhysReflect.

SubsetCasesOnline preference95% CIPairwise agreement
All20068.8%62.40–75.20%74.49%
Real (MSD + CC-Mirror)10083.4%80.85–86.15%82.90%

The study includes 50 single-target, 50 multi-target, 50 MSD, and 50 CC-Mirror cases. Participants assessed reflected-object position, scale, orientation, occlusion, and source–reflection correspondence. They were recruited through university advertisements, excluding authors, their students, and research collaborators. Method identities were hidden, presentation order randomized, and GT reflections withheld. All participants completed the full evaluation without missing or invalid responses.

Confidence intervals use 20,000 category-stratified two-way bootstrap resamples of participants and images (Owen, 2007). Pairwise agreement measures the proportion of participant pairs selecting the same method for each image.

The original six-method study separately involved five participants evaluating 200 cases across three perceptual criteria, yielding 3,000 valid selections.

Metric discrepancy. As illustrated in Appendix B.3, IoU/BF1 measure reference-object alignment, whereas GLAM-GEO assesses broader spatial coherence and may overlook object displacement. Closer reflected-object alignment can coexist with inconsistencies in surrounding scene structure; these metrics therefore capture different aspects of reflection quality and are interpreted separately.

We clarify PhysReflect's cross-group advantages in Table 1 and correct Appendix Table 5, where some entries inadvertently retained single-object results during appendix assembly. The expanded table also includes GPT Image 2 and MirrorRefine-online.

AuthorsOpenReview ↗

W2 / A1 / A2 / A3-1: Agent-loop contribution and computational cost

We provide complete standalone Fill results, controlled same-backend comparisons against simpler alternatives, and measured computational costs, demonstrating quality and efficiency benefits of the integrated reflection-guided refinement procedure.

Standalone and controlled comparisons (A1–A2). All experiments use the same 500 MirrorBenchV2 samples. Independent best-of-N generates Fill candidates and selects the highest-scoring output using the Qwen3.8-27B evaluator (Geometry/Lighting/Appearance weights: 40%/30%/30%). The same weighted score is also used in the stopping procedure of MirrorRefine-local and the matched iterative controls. Candidate retention follows the verification rule described in Appendix A.3. Fixed-mask feedback performs iterative VLM critique, prompt revision, and full-mirror inpainting, retaining the same reflection context, RSG, candidate verification, retention, and stopping conditions as all-Fill. The same-backend controls share FLUX.1-Fill-dev and an eight-attempt synthesis cap.

MethodPSNR ↑SSIM ↑LPIPS ↓CLIP ↑IoU ↑BF1 ↑Geo ↑Light ↑App ↑Attempts
Standalone Fill13.44650.74870.202126.66500.40240.146856.5569.7868.831.00
Fill best-of-816.03040.74820.171326.04340.41250.160768.6571.1671.658.00
Fill best-of-415.80140.74470.172625.81670.40980.159868.3370.8771.414.00
Fixed-mask feedback16.21210.74890.168825.90330.41370.161068.7771.4272.056.34
MirrorRefine all-Fill16.23650.75080.163226.18100.42690.162169.0871.8372.304.24
Full MirrorRefine-local16.53890.75380.164126.17560.42730.167369.5472.7375.664.06

All-Fill improves eight of nine metrics over standalone Fill, increasing PSNR by 2.79 dB, IoU by 0.0245, BF1 by 0.0153, and Geometry by 12.53 points, while reducing LPIPS by 0.0389. Standalone Fill retains higher CLIPScore.

All-Fill also improves all nine metrics over best-of-8 and fixed-mask feedback. Compared with best-of-8, IoU increases by 0.0144 and LPIPS decreases by 0.0081, using 47.0% fewer synthesis attempts. Relative to fixed-mask feedback, IoU increases by 0.0132, while attempts decrease from 6.34 to 4.24.

These comparisons support the integrated iterative procedure over independent selection and adaptive repair localization over fixed-mask feedback. Full additionally uses FLUX.1-Kontext-dev for targeted editing; all-Fill provides the strictly same-backend comparison.

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

This paper presents a multi-agent iterative refinement framework for synthesizing mirror reflections. Given a prompt and an optional source image with a mirror mask, a Reflection Optimization Controller (Qwen3.8-27B locally or GPT-5.6 Sol online) describes visible exterior objects and performs Source–Mirror–Reflection reasoning to infer the expected reflected appearance and viewpoint. A Quality Evaluator combines an offline-trained Reflection Spatial Geometry (RSG) module with object detection and a VLM judge to diagnose geometric, appearance, and lighting defects. Across up to 8 synthesis calls, the evaluator selects among full-image generation (Action G), mirror in-painting (Action I), or targeted editing (Action E), retaining candidates only when targeted defects are resolved without introducing new errors. Experiments evaluate MirrorRefine-local (FLUX tools + Qwen3.8-27B) and MirrorRefine-Online (GPT Image 2 + GPT-5.6 Sol) on fixed-scene mirror completion across MirrorBenchV2, MSD, and CC-Mirror.

Critical strengths

The paper is well motivated. I think that the the main observation is that existing prompt-refinement and image-editing agents fail on mirror scenes because they optimize global text–image alignment rather than reasoning about mirror reflection geometry. The ablation study on the local pipeline (Table 3) shows that providing Source–Mirror–Reflection context and coarse spatial priors from RSG substantially improves FLUX completion on MirrorBenchV2 (raising PSNR from 13.67 to 16.54 and reflection-mask IoU from 0.3895 to 0.4273 while reducing average rounds from 5.4 to 4.1), and generalizes to real-world scenes on MSD and CC-Mirror without fine-tuning the diffusion generator.

Critical weaknesses

I have four main concerns regarding the experimental evaluation and technical specification:

  1. Failure to beat 2026 SOTA (PhysReflect) on objective metrics and judge–controller circularity in GLAM-Bench (Tables 1, 4, and 5): On MirrorBenchV2—the only benchmark with ground-truth 3D reflections and object masks—the 2026 baseline PhysReflect (Ge et al., Sept 2026, budget 4) substantially outperforms both MirrorRefine-local and MirrorRefine-Online (budget 8) on every objective reference and geometric metric: PSNR (19.12 vs. 16.54 / 16.65), SSIM (0.7731 vs. 0.7538 / 0.7459), LPIPS (0.1071 vs. 0.1641 / 0.1660), reflection-mask IoU (0.6932 vs. 0.4273 / 0.4069), Boundary IoU (0.5287 vs. 0.2382), and Boundary F1 (0.4228 vs. 0.1673 / 0.1477). In fact, MirrorRefine-Online's ground-truth IoU (0.4069) is lower than single-step unrefined Lumina-Image 2.0 (0.4088), and MirrorRefine-Online is omitted entirely from Table 5. Despite much worse ground-truth geometry (IoU 0.4069 vs. 0.6932), MirrorRefine-Online claims the highest Geometry score on GLAM-Bench (74.43 vs. 68.10). However, Appendix B.1 reveals that GLAM-Bench uses GPT-5.6 Sol as its judge—the exact same VLM that MirrorRefine-Online uses as both its Controller and Evaluator. Comparing "Online / GPT" in Table 4 (Qwen3.8-27B controller + GPT-5.6 Sol evaluator + GPT Image 2) with MirrorRefine-Online in Table 1 (GPT-5.6 Sol controller + GPT-5.6 Sol evaluator + GPT Image 2) exposes this circularity: switching the controller to GPT-5.6 Sol causes ground-truth geometric IoU to drop from 0.4239 to 0.4069, while GLAM-Bench Geometry (judged by GPT-5.6 Sol) jumps by +5.01 points from 69.42 to 74.43.

  2. Omission of five 2026 baselines cited in Section 2 and missing single-call baseline for GPT Image 2 (2026):

    • Section 2 discusses five directly relevant 2026 methods: Fill My Mirror (Basson et al., Sept 2026), PhysMirror (Mai et al., July 2026), M3 (Yang et al., Feb 2026), CRAFT (Kovalev et al., Jan 2026), and VisualPrompter (Wu et al., ICLR 2026), yet none of these five are included in Tables 1, 5, 6, or 7. Also, for ReflectionFlow (ICCV 2025), Appendix B.2 notes that the authors did not run the reflection-tuned model.
    • While MirrorRefine-Online uses the 2026 GPT Image 2 model as its synthesis backbone, a single-pass (budget 1) GPT Image 2 baseline is never reported in Tables 1, 5, 6, or 7, making it impossible to separate the gain of the agentic loop from the base generator.
  3. Disconnect between the three-action framework (Action G, Action I, Action E) and the experiments: All experiments only evaluate fixed-scene mirror completion where full-image generation (Action G) is disabled. Moreover, between Action I (inpainting) and Action E (editing), Table 3 shows that "Fill for all actions"—which disables Action E and calls FLUX.1-Fill-dev in every round—achieves virtually identical IoU (0.4269 vs. 0.4273), better LPIPS (0.1632 vs. 0.1641), and better CLIPscore (26.1810 vs. 26.1756). Similarly, for MirrorRefine-Online, GPT Image 2 does not accept a binary mask, so Action I and Action E both reduce to instruction-based calls to the same model.

  4. Underspecified RSG, Object Detection, and Layout Judge: Section 3.3 and Appendix A.2 omit the training loss of SegFormer-B0, which object detector is used in Object Detection, and how Layout Judge matches multiple exterior objects to RSG's single class-agnostic binary mask when reflected objects merge into one connected region.

Questions for authors
  1. Independent VLM judge for GLAM-Bench: Why does switching the controller from Qwen3.8-27B (Table 4) to GPT-5.6 Sol (Table 1) reduce ground-truth geometric IoU from 0.4239 to 0.4069 while increasing GLAM-Bench Geometry (judged by GPT-5.6 Sol) from 69.42 to 74.43? Can you report GLAM-Bench using an independent third-party VLM judge and include MirrorRefine-Online in Table 5?

  2. Missing 2026 baselines: What are the metrics of single-pass (budget 1) GPT Image 2 across Tables 1, 5, 6, and 7, and why were Fill My Mirror, PhysMirror, M3, CRAFT, and VisualPrompter (all cited in Section 2) omitted from the experimental tables?

  3. Justification of multi-action switching and Action G: Can you provide quantitative results for text-to-image generation (Action G) and explain why switching between Action I and Action E is needed when "Fill for all actions" in Table 3 performs equally well?

  4. Implementation details of RSG and Layout Judge: What loss function trains RSG, which object detector is used, and how is Layout Judge implemented?

8 comments
AuthorsOpenReview ↗

W1: Reference-based performance relative to specialized mirror-generation methods

PhysReflect’s stronger IoU/BF1 results are obtained with mirror-task training on SynMirrorV2, whose held-out scenes form MirrorBenchV2, and rendered ground-truth depth outside the mirror as an additional input during generation.

Like the evaluated inference-time scaling baselines, MirrorRefine uses frozen pretrained generative backbones in a training-free manner, without mirror-task training or fine-tuning of these backbones. For fixed-scene completion, it takes a single mirror-erased RGB image and a mirror mask as input, without depth maps. Its RSG module is trained offline on the separate Mirror-Reflect-Inference dataset and remains frozen during refinement (Appendix A.2).

We additionally report MirrorVerse and PhysReflect configurations without depth input. We remove the depth-map input, retain the remaining model architecture, and retrain each model following the settings specified in its respective paper:

MethodIoU ↑BF1 ↑GLAM-GEO ↑
MirrorVerse without depth0.41640.131560.14
PhysReflect without depth0.46110.139361.05
MirrorRefine-local0.42730.167369.54
MirrorRefine-Online0.40690.147774.43

PhysReflect’s IoU/BF1 change from 0.6932/0.4228 with depth to 0.4611/0.1393 after retraining without depth input, narrowing its IoU gap to Online from 0.2863 to 0.0542. Both MirrorRefine configurations achieve higher BF1 and GLAM-GEO than these retrained baselines, while PhysReflect without depth retains the highest IoU.

Online’s main IoU (0.4069) is also slightly below Lumina-Image 2.0 (0.4088); the comparison with its own synthesis backbone is presented in W2/Q2 below.

AuthorsOpenReview ↗

W1/Q1: Additional VLM judges and differences between IoU/BF1 and GLAM-GEO

On the same 500 MirrorBenchV2 samples used in the main experiments, both additional VLM judges consistently rank MirrorRefine-Online above the Qwen-controller hybrid and PhysReflect across Geometry, Lighting, and Appearance.

Controller comparison and additional judges. The Table 4 hybrid uses a Qwen3.8-27B Controller, whereas full Online uses GPT-5.6 Sol for control. Both use the same GPT-5.6 Sol Quality Evaluator and GPT Image 2 synthesis backend, so only the Controller changes in this comparison. To examine dependence on the original GPT-5.6 Sol judge, we re-score the same method-specific final outputs with the third-party DeepSeek-V4.1-Flash-1017 judge and an additional GPT-6.1 Sol judge. The results are below:

JudgeMethodGeometry ↑Lighting ↑Appearance ↑
DeepSeek-V4.1-Flash-1017PhysReflect61.8879.6677.88
DeepSeek-V4.1-Flash-1017Qwen-controller hybrid66.8584.0286.35
DeepSeek-V4.1-Flash-1017MirrorRefine-Online68.2384.7988.78
GPT-6.1 SolPhysReflect81.0286.3179.02
GPT-6.1 SolQwen-controller hybrid82.3388.1787.54
GPT-6.1 SolMirrorRefine-Online85.1889.6890.68

Under DeepSeek, Online improves over the Qwen-controller hybrid by 1.38 points in Geometry, 0.77 in Lighting, and 2.43 in Appearance. Under GPT-6.1 Sol, the corresponding improvements are 2.85, 1.51, and 3.14 points. Both additional judges therefore preserve the same ordering across all three evaluation dimensions.

In the original evaluation, changing the Controller decreases IoU from 0.4239 to 0.4069, while increasing GLAM-GEO from 69.42 to 74.43. IoU measures agreement with annotated reflected-object masks, whereas GLAM-GEO assesses perceived spatial coherence across the reflected scene. When constructing and revising prompts, the GPT-based Controller may place greater emphasis on the overall spatial layout, background structure, and inter-object relationships within the reflected scene. These aspects are not fully captured by reference-mask overlap, so higher GLAM-GEO can coexist with lower IoU. This is a plausible interpretation of the observed differences, rather than an established causal explanation.

We have also included Online and single-pass GPT Image 2 in the complementary geometry comparison in Table 5:

MethodBIoU ↑IoU_obj ↑BF1_obj ↑
GPT Image 2, B = 10.18660.33130.1358
PhysReflect0.52870.63970.4228
MirrorRefine-local0.23820.37490.1532
MirrorRefine-Online0.22170.37260.1479

The analysis of differences between IoU/BF1 and GLAM-GEO metrics, together with illustrative examples, is provided in Appendix B.3.

AuthorsOpenReview ↗

W2/Q2: Additional baselines and single-pass GPT Image 2

We have added results for Fill My Mirror, PhysMirror, M3, CRAFT, VisualPrompter, and the full ReflectionFlow system, as reported below. The ReflectionFlow row evaluates its reflection-tuned model and corresponding inference pipeline, rather than the Noise & Prompt Scaling baseline, which adapts only the noise- and prompt-level mechanisms from Zhuo et al. (2025). Single-pass GPT Image 2 is also included, with results reported in Tables 1, 5, 6, and 7.

The added comparisons use the officially released models and weights. The synthesis-attempt caps are B = 4 for Fill My Mirror and PhysMirror, B = 8 for ReflectionFlow, M3, CRAFT, and both MirrorRefine configurations, and B = 2 for VisualPrompter. Single-pass GPT Image 2 uses B = 1. Bold indicates the best value within this supplementary MirrorBenchV2 table.

MethodPSNR ↑SSIM ↑LPIPS ↓CLIPScore ↑IoU ↑BF1 ↑Geo ↑Light ↑App ↑
GPT Image 2, B = 114.36180.70090.199026.49470.38260.136366.9970.1072.07
Fill My Mirror16.84890.75880.171126.21280.39980.159766.5671.0873.71
PhysMirror11.38030.76530.242625.58310.38950.135048.9170.4062.09
ReflectionFlow13.98790.74260.217326.38990.40440.140763.6367.7366.57
M312.45010.73740.238626.07800.39570.137259.1767.1069.38
CRAFT12.44970.73310.243326.36640.40410.138362.9369.7469.04
VisualPrompter12.80920.74560.235626.42660.39230.129259.3971.4463.44
MirrorRefine-local16.53890.75380.164126.17560.42730.167369.5472.7375.66
MirrorRefine-Online16.65130.74590.166026.79140.40690.147774.4374.8478.74

Online improves all nine metrics over ReflectionFlow, M3, CRAFT, and VisualPrompter. Local achieves lower LPIPS and higher IoU, BF1, and all three GLAM scores than the added baselines. Fill My Mirror has the highest PSNR in this table, and PhysMirror has the highest SSIM.

Single-pass GPT Image 2. Online improves all nine main metrics on MirrorBenchV2, as shown above, and all three complementary geometry metrics reported in W1/Q1. The corresponding MSD and CC-Mirror results are:

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓CLIPScore ↑Geo ↑Light ↑App ↑
MSDGPT Image 2, B = 116.72830.84100.133728.800373.1871.0177.75
MSDMirrorRefine-Online17.590.86210.121129.4881.9075.8788.08
CC-MirrorGPT Image 2, B = 117.48590.79630.135826.783665.0372.9678.35
CC-MirrorMirrorRefine-Online18.660.81990.119726.3678.1675.2683.07

Online improves all seven metrics on MSD and six of seven on CC-Mirror; the exception is CC-Mirror CLIPScore, where single-pass GPT Image 2 remains higher (26.7836 versus 26.36). The same-backbone comparison establishes gains of the complete refinement system under the stated budgets, but does not separate individual refinement components from the additional synthesis attempts. The broader baseline comparisons also involve different backbones and refinement procedures.

The additional baseline results are reported in Appendix B.3. The single-pass GPT Image 2 results are included in Tables 1, 5, 6, and 7.

Reviewer_9WrFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MirrorRefine improves mirror-region completion through iterative diagnosis and repair. A controller infers reflected content from the source scene. An offline-trained Reflection Spatial Geometry module predicts coarse layout regions, which a vision-language evaluator uses alongside appearance and lighting checks. The system selects inpainting or editing actions and retains candidates judged to repair targeted defects while preserving other content. Local and hosted configurations are evaluated on MirrorBenchV2, MSD and CC-Mirror, with component ablations and a human preference study.

Critical strengths
  1. The spatial and semantic cues are tied to the task and tested within one pipeline. Removing RSG reduces local IoU from 0.4273 to 0.3967 and GLAM geometry from 69.54 to 63.37 (Table 3). The separate context and joint-removal ablations help show that both sources of information contribute. The system also distinguishes uncertain observations from verified repairs rather than assuming every candidate improves the image.

  2. The local configuration has supporting evidence beyond its automated evaluator. It receives 42.77% of best-image selections in the six-method user study, versus 31.37% for PhysReflect (Table 2), and performs well on the listed real-image completion metrics. These preferences do not certify geometric correctness, but they support practical perceptual value. The paper also reports failure cases and trade-offs between reference-image and evaluator-based metrics.

Critical weaknesses
  1. The distinct contribution of adaptive action selection and candidate retention is not isolated. The main comparisons change generators, selectors and synthesis budgets. Table 3's all-Fill variant is already close to the full system in IoU (0.4269 versus 0.4273) and slightly better in LPIPS (0.1632 versus 0.1641). The context/RSG ablations are useful, but they do not establish the benefit of the adaptive loop over simple reranking or fixed-action refinement using the same backend, spatial prior, evaluator and attempt cap. This matters for assessing the contribution beyond assembling capable existing tools.

  2. The evidence for reflection geometry is mixed and partly dependent on the evaluator used during refinement. On MirrorBenchV2, the local and online IoUs are 0.4273 and 0.4069, versus 0.6932 for PhysReflect, while the online system receives the highest GLAM geometry score (Table 1). These metrics need not rank all valid reflections identically, but the discrepancy warrants investigation. Appendix B.1 acknowledges that separate scoring contexts do not remove shared GPT-5.6 Sol biases; the human study evaluates only the local configuration. Indeterminate AI judgments are also excluded on the real-image datasets without reported per-method counts. This limits confidence that the online geometry advantage reflects better reflection consistency.

Questions for authors
  1. At the same eight-attempt cap, how does the full local pipeline compare with reranking independently generated Fill candidates using the same RSG and evaluator, or with fixed-action refinement? One focused control would help isolate what adaptive action selection and retention contribute.

  2. Can you analyze the existing outputs with high GLAM scores but poor geometric overlap, ideally with blinded human judgments of reflection consistency? Please include the online configuration and report the number of valid versus indeterminate AI judgments per method on MSD and CC-Mirror. This would clarify whether the metric disagreement reflects valid alternative reflections, evaluator error or other factors.

4 comments
AuthorsOpenReview ↗

W1: The contribution of adaptive action selection and candidate retention is insufficiently isolated, and all-Fill performs similarly to Full.

We thank the reviewer for recognizing the complementary contributions of reflection-context reasoning and RSG, as well as the improvements in overall image quality supported by the local results.

The all-Fill variant retains our adaptive refinement loop, and the new same-backend controls support the contribution of this integrated procedure. The full system provides additional boundary and appearance benefits, with trade-offs across metrics.

The all-Fill variant retains reflection-context reasoning, RSG-guided diagnosis, adaptive inpainting regions, and candidate verification, while using FLUX.1-Fill-dev for every synthesis action. The all-Fill and Full variants therefore compare different synthesis configurations within the refinement procedure. Relative to all-Fill, Full improves PSNR from 16.2365 to 16.5389, BF1 from 0.1621 to 0.1673, and Appearance from 72.30 to 75.66, while IoU remains nearly unchanged (0.4269 versus 0.4273). All-Fill retains slightly better LPIPS and CLIPScore. We interpret the additional editing capability as providing complementary benefits, particularly in boundaries and appearance.

To assess the contribution beyond these synthesis choices, we provide controlled comparisons with independent best-of-N selection and fixed-action feedback refinement under Q1 below. With the Fill backend held fixed, all-Fill improves all nine reported quality metrics over both alternatives. RSG-guided diagnosis identifies the error, repair region, and intended correction; the controller updates the synthesis conditions, and candidate verification determines whether to retain the resulting repair. Figure 5 illustrates this progression from initial completion to targeted correction. The comparison with independent best-of-N selection evaluates the integrated iterative diagnosis-and-repair procedure, while the fixed-mask feedback comparison tests adaptive repair-region selection with the backend, candidate retention, and stopping conditions held fixed. Candidate retention is shared by the iterative variants.

AuthorsOpenReview ↗

Q1: At the same eight-attempt cap, how does the full local pipeline compare with independent best-of-N Fill selection and fixed-action refinement?

At the same eight-attempt cap, both all-Fill and Full outperform best-of-8 selection and generic-feedback refinement on all nine quality metrics. All-Fill averages 4.2 rounds and Full 4.1, compared with 8.0 for best-of-8 and 6.3 for generic feedback. The all-Fill comparison establishes the advantage with the generation backend held fixed.

All configurations use the same RSG and local Quality Evaluator (Qwen3.8-27B) and include the same reflection context. Independent best-of-N selection generates N Fill candidates independently (N = 8 for best-of-8 and N = 4 for best-of-4), evaluates each candidate with the same Quality Evaluator used by MirrorRefine-local, and returns the single candidate with the highest quality score. This score is a weighted sum of the evaluator’s geometry, lighting, and appearance scores, using the same weights as MirrorRefine-local: 40%, 30%, and 30%, respectively. The same weighted score is also used in the stopping procedure of MirrorRefine-local and the matched iterative controls. Candidate retention follows the verification rule described in Appendix A.3. The “Generic feedback + Fill” baseline repeatedly performs VLM critique, prompt revision, and Fill completion with a fixed full-mirror mask. It retains candidate verification and retention, starts each refinement step from the retained candidate, and receives the same reflection context and RSG information as all-Fill. It also follows the same stopping conditions. The fixed full-mirror mask is the only difference from all-Fill; no other mechanism is removed.

All-Fill uses the same Fill backend with evaluator-determined adaptive inpainting regions, while Full additionally employs FLUX.1-Kontext-dev for Action E. The all-Fill variant corresponds to “Fill for all actions” in Table 3. The controls use the same MirrorBenchV2 samples and other experimental settings as the original ablations in Appendix B.5. The main comparisons use an eight-attempt cap, including initialization, failures, and retries; best-of-4 provides a supplementary comparison near the iterative methods’ average generation rounds.

MethodPSNR ↑SSIM ↑LPIPS ↓CLIPScore ↑IoU ↑BF1 ↑Geo ↑Light ↑App ↑Avg. rounds
Independent Fill, best-of-816.03040.74820.171326.04340.41250.160768.6571.1671.658.0
Independent Fill, best-of-415.80140.74470.172625.81670.40980.159868.3370.8771.414.0
Generic feedback + Fill (fixed mask)16.21210.74890.168825.90330.41370.161068.7771.4272.056.3
MirrorRefine all-Fill16.23650.75080.163226.18100.42690.162169.0871.8372.304.2
Full MirrorRefine-local16.53890.75380.164126.17560.42730.167369.5472.7375.664.1

Compared with best-of-8, all-Fill improves IoU from 0.4125 to 0.4269, BF1 from 0.1607 to 0.1621, and Geometry from 68.65 to 69.08, while reducing LPIPS from 0.1713 to 0.1632. These gains are achieved with 47.5% fewer generation rounds on average (4.2 versus 8.0). Increasing independent sampling to eight candidates therefore does not close the observed quality gap under the tested settings.

All-Fill also improves all nine quality metrics over the fixed-mask generic-feedback control, including IoU from 0.4137 to 0.4269 and LPIPS from 0.1688 to 0.1632, while reducing average rounds from 6.3 to 4.2. With candidate retention and stopping conditions held fixed, this comparison supports the contribution of adaptive repair-region selection. Against best-of-4, it improves all nine metrics at comparable average rounds (4.2 versus 4.0), including IoU from 0.4098 to 0.4269.

Full also outperforms best-of-8 across all nine quality metrics with 4.1 versus 8.0 average rounds, including IoU (0.4273 versus 0.4125), BF1 (0.1673 versus 0.1607), and Appearance (75.66 versus 71.65). These controls support improvements in quality and average generation rounds over independent best-of-N selection and generic-feedback refinement.

AuthorsOpenReview ↗

W2: The evidence for reflection geometry is mixed, and Online’s GLAM advantage may depend on the evaluator used during refinement.

The stronger IoU/BF1 results of the mirror-specific baseline PhysReflect are obtained with mirror-task training on SynMirrorV2, whose held-out scenes form MirrorBenchV2, and rendered ground-truth depth outside the mirror as an additional input during generation. Like the evaluated inference-time scaling baselines, MirrorRefine uses frozen pretrained generative backbones in a training-free manner, without mirror-task training or finetuning, and takes a single mirror-erased RGB image and a mirror mask as input, without depth maps. Its RSG is trained offline on the separate Mirror-Reflect-Inference dataset and remains frozen during refinement (Appendix A.2).

The additional comparisons without depth substantially narrow PhysReflect’s IoU advantage and reverse its BF1 advantage over both MirrorRefine configurations. We evaluated MirrorVerse and PhysReflect without depth inputs to examine the role of this additional modality:

MethodIoU ↑BF1 ↑GLAM-GEO ↑
MirrorVerse without depth0.41640.131560.14
PhysReflect without depth0.46110.139361.05
MirrorRefine-local0.42730.167369.54
MirrorRefine-online0.40690.147774.43

PhysReflect’s IoU/BF1 decrease from 0.6932/0.4228 with depth to 0.4611/0.1393 without depth, reducing its IoU gap relative to Online from 0.2863 to 0.0542. MirrorRefine-online achieves a higher BF1 than PhysReflect without depth input (0.1477 versus 0.1393). Both MirrorRefine configurations also obtain higher GLAM-GEO scores than the two baselines without depth. PhysReflect without depth retains higher IoU than both MirrorRefine configurations, and Online’s IoU remains below MirrorVerse without depth. These additional comparisons complement the original results with depth and show the influence of depth availability on reference alignment.

We separately examine whether Online’s GLAM advantage depends on the evaluation model. Additional evaluations use DeepSeek-V4.1-Flash-1017 and GPT-6.1 Sol, comparing Online with PhysReflect and RAISE:

JudgeMethodGeometry ↑Lighting ↑Appearance ↑
DeepSeek-V4.1-Flash-1017PhysReflect61.8879.6677.88
DeepSeek-V4.1-Flash-1017RAISE29.8768.0137.70
DeepSeek-V4.1-Flash-1017MirrorRefine-online68.2384.7988.78
GPT-6.1 SolPhysReflect81.0286.3179.02
GPT-6.1 SolRAISE40.2080.8045.32
GPT-6.1 SolMirrorRefine-online85.1889.6890.68

Both additional judges rank Online above PhysReflect and RAISE in all three dimensions, with Geometry margins of 6.35 and 4.16 points over PhysReflect, respectively. This consistency reduces the concern that the observed GLAM advantage is specific to the original GPT-5.6 Sol judge.

To assess whether Online’s benefits extend beyond VLM scores, we additionally evaluated reflected-object alignment on MSD and CC-Mirror. Ground-truth target masks were constructed by applying SAM3 semantic segmentation to the reference images, excluding segmented objects that were neither mentioned in the text nor visible outside the mirror, and manually verifying and correcting the retained masks. We followed the MirrorBench evaluation protocol for IoU and BF1, with the BF1 boundary-matching tolerance relaxed to 2 pixels (BF1@2px), and used the same implementation for all methods.

MethodMSD IoU ↑MSD BF1@2px ↑CC-Mirror IoU ↑CC-Mirror BF1@2px ↑
PhysReflect0.11670.03010.08050.0375
PhysReflect (finetuned on MSD)0.18270.0477Not testedNot tested
RAISE0.14800.04180.13800.0484
MirrorRefine-online0.21510.05440.24410.0771

Online achieves higher IoU and BF1@2px than the evaluated baselines on both datasets. On MSD, it also outperforms PhysReflect finetuned on MSD. These results provide reference-based measurements of object regions and boundaries on real images, complementing the evaluation of the overall scene. The Online examples and human study under Q2 provide further qualitative and human-preference evidence.

AuthorsOpenReview ↗

Q2: What explains the GLAM-GEO/IoU–BF1 discrepancy, how many AI judgments are valid, and what does the blinded Online evaluation show?

We have added representative examples of this metric discrepancy in Figure 7 of Appendix B.3 (“Additional quantitative comparisons”) in the revised manuscript. These examples show that closer alignment of the annotated reflected objects can coexist with discrepancies in the surrounding scene. In our evaluation, IoU/BF1 measure agreement with the GT regions and boundaries of the annotated targets, so they do not directly assess scene structure outside those annotations. In the second row of Figure 7, for example, PhysReflect aligns the reflected container more closely with its pink GT contour, while the green-boxed region shows its reflected tabletop extending into the area occupied by the vertical background in GT. Online preserves the tabletop–background separation, but its reflected container is displaced relative to the GT contour. This displacement is a reference-alignment error; the higher GLAM-GEO score does not negate it. The green boxes in the first and third rows highlight background discrepancies relative to GT. Such differences in scene structure outside the annotated target regions may influence perceived spatial coherence. GLAM-GEO is intended to assess spatial coherence of the overall reflected scene using the source context and the complete generated image. We believe that these differences in evaluation coverage explain part of the divergent rankings; limited sensitivity of the VLM judge to object-placement errors may also contribute. A more detailed analysis of these examples and the metric discrepancy is provided in Appendix B.3.

For every compared method and each of the three GLAM dimensions, the original valid/unscored counts are identical: 200/0 on MSD and 199/1 on CC-Mirror. The final common evaluation sets contain 200 MSD images and 199 CC-Mirror images. These are the original per-method, per-dimension counts, not only the sizes of the common sets after exclusion.

DatasetOriginal valid images per method and dimensionOriginal unscored images per method and dimensionReason
MSD200 of 2000None
CC-Mirror199 of 2001Evaluator content-policy refusal

The unscored CC-Mirror case resulted from an evaluator content-policy refusal and was excluded from the common comparison set. Reported mean AI scores are computed over the same valid image set for all compared methods and dimensions; the unscored case is excluded rather than assigned zero.

MirrorRefine-online was preferred over PhysReflect for geometric plausibility in a blinded pairwise study involving 15 volunteers and 200 test images. Each volunteer evaluated all 200 images, yielding 3,000 valid geometric-plausibility judgments. Each judgment was a forced choice between the two outputs, with no tie option. The reported 68.8% is the proportion of all valid geometric-plausibility votes assigned to Online. The table reports geometric-plausibility preferences for ALL test images and, as a separate criterion, image-quality preferences for the combined real-world subset from MSD and CC-Mirror:

MethodALL (geometric plausibility)Real world (geometric plausibility)
PhysReflect31.2%16.6%
MirrorRefine-online68.8%83.4%

The evaluation set comprised 50 single-target cases, 50 multi-target cases, 50 MSD images, and 50 CC-Mirror images. ALL contains 200 images, and the combined real-world subset contains 100 images. Test images were presented in random order, and the ordering of the two outputs within each comparison was also randomized. The compared outputs were displayed without method identifiers, and assessments were collected anonymously. Participants were shown the source image, original prompt, and mirror region alongside the two outputs; the GT reflection was not shown. For the geometric-plausibility evaluation, participants were explicitly instructed to assess object position, scale, orientation, occlusion, and correspondence between source objects and their reflections, and to select the result with more plausible reflection geometry.

Across the evaluated 200-image set, the blinded study shows that participants favor Online over PhysReflect when judging the geometric plausibility of the generated reflections, complementing the reference-based alignment measurements and the evaluations with additional GLAM judges. The real-world results also show a preference for Online in overall geometric plausibility.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.