acceptodds
Under review as a conference paper at ICLR 2027

VIEScore2: Unified Image Evaluation with Spatially Grounded Explanations

Abstract

Existing synthetic image evaluators typically provide only a scalar quality score and do not identify the image regions that support it. We introduce VIEScore2, a unified evaluator for image generation and editing tasks with optional conditioning images. VIEScore2 represents an image as an NxN grid and jointly predicts quality scores and defect locations in a single model pass. Its text-native grid representation provides a common interface for heterogeneous spatial supervision and enables directly verifiable post-training objectives. We train on 38K examples spanning score-only, localization-only, and joint supervision across generation and editing tasks. Starting from supervised fine-tuning, we further apply GRPO to improve defect localization using rewards that combine cell-level Dice overlap, score accuracy, and output-format validity. A parameter-free parser converts the structured predictions into readable explanations. On the primary suite, VIEScore2 achieves an overall-score SRCC of 0.601, compared with 0.491 for Gemini-3-Flash, the strongest zero-shot general-purpose VLM baseline under matched inputs. For defect localization, VIEScore2 outperforms both general-purpose VLMs and specialized spatial evaluators on three of six benchmarks in per-image grid IoU and ranks among the top three on five, including datasets beyond its training sources.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Synthetic Image Evaluation, Defect Localization, Vision-Language Models

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

This paper presents VIEScore2, a unified evaluator for generated and edited images. The model predicts perceptual quality and semantic consistency scores, as well as spatial locations of artifacts and prompt–image misalignment. The authors convert heterogeneous localization annotations into a common 16 × 16 grid format and train the model with supervised fine-tuning followed by GRPO. They evaluate its agreement with human judgments and its defect localization performance across several benchmarks.

Critical strengths
  1. The joint treatment of overall scoring and defect localization is useful. The localized outputs also make the scores easier to inspect.
  2. The common grid format offers a practical way to train on datasets with different types of spatial annotations.
Critical weaknesses
  1. The paper presents VIEScore2 as a potential reward model, but does not directly test whether using it as a reward improves a generative model.
  2. The results do not yet make clear how well its scoring and localization generalize to the defects produced by recent generators. Although SDG-30K includes newer models, a generator-level breakdown would be helpful.
Questions for authors
  1. Since reward modeling is an important potential use of VIEScore2, could the authors conduct a small-scale experiment using it as the reward in one training method, such as Flow-GRPO, DanceGRPO, or DiffusionNFT? The key question is whether optimizing with VIEScore2 improves generated images under an independent evaluation. If feasible, a comparison with an established reward-model baseline under the same generator and training budget would make the result easier to interpret, but even a limited experiment would be informative.

  2. Many images in the training datasets come from earlier generators, and Table 14 largely focuses on older models. As generation and editing models improve, the types and visibility of artifacts may change. Could the authors report results separately for the newer generators already represented in SDG-30K and, if feasible, conduct a small additional evaluation on models such as FLUX.2 [klein] or Z-Image? I see this as a generalization question rather than a requirement for a large new benchmark.

  3. I found two examples in Figure 3 difficult to interpret. In the SynthScars example, the annotated region appears to cover only part of the unclear text, whereas the GPT-5.6-sol prediction covers more of the text. In the SDG-30K example, the stated asymmetry in the porcelain figure’s facial gold inlays is hard to discern even when enlarged. Could the authors show the original annotations for these cases and check whether similar annotation ambiguities occur elsewhere in the training or test data?

5 comments

We appreciated the reviewer's concrete suggestions and ran each of the requested experiments.

New samples from recent generators (used in Q1 and Q2). We generated 8 images for each of the 200 DrawBench prompts with Z-Image-Turbo and FLUX.2-klein-4B, and scored every image with VIEScore2, PickScore, and HPSv2.1. For each prompt, we compare the image VIEScore2 scores highest, or lowest, with a random one of the 8:

Mean PickScore of the chosen imageZ-Image-TurboFLUX.2-klein-4B
Random candidate22.8722.83
Highest VIEScore2 score22.9822.95
Lowest VIEScore2 score22.7222.68

On both generators, neither of which was seen in training, the image VIEScore2 scores highest has a higher PickScore than a random one, and the image it scores lowest has a lower one. On these recent generators, VIEScore2’s ordering therefore agrees with an independent preference model.

VIEScore2 as a reward signal (W1, Q1).

Using VIEScore2 to train generators lies beyond the scope of the submission, which focuses on evaluation. Following the reviewer’s suggestion, we nevertheless ran two small-scale experiments, and both results are encouraging.

(a) Best-of-N selection. On the samples above, choosing the image VIEScore2 scores highest improves PickScore over random selection on both generators. Within a prompt, VIEScore2 correlates only weakly with PickScore and HPSv2.1 (Spearman 0.09–0.16), so it adds a complementary signal.

(b) RL fine-tuning (new experiment). We fine-tuned FLUX.2-klein-base-4B with DiffusionNFT (LoRA, 160 steps), using VIEScore2’s overall score or PickScore as the reward under the same training budget, and evaluated on DrawBench with metrics that neither reward uses. As in DiffusionNFT, all models are sampled without CFG, which is why the base model has a negative ImageReward.

DrawBench (200 prompts)HPSv2.1ImageRewardCLIPScoreAesthetic
FLUX.2-klein (base)0.204−0.560.2374.94
+ PickScore reward0.2720.750.2486.09
+ VIEScore2 reward0.2770.810.2526.13

VIEScore2 improves every metric over the base model and matches or slightly exceeds PickScore on all four. This includes HPSv2.1 and ImageReward, which are preference models of the same kind as PickScore.

Recent generators (W2, Q2).

VIEScore2's grid IoU on SDG-30K by generator (same protocol as Table 1):

Grid IoUSANA-1.5FLUX.2Z-ImageLongCat-ImageAll four (paper)
VIEScore20.1120.0860.1010.1020.102 (Table 1)
VIEScore2 without GRPO0.0760.0420.0340.0490.052 (Table 17)
ImageDoctor0.0930.0840.0880.0780.086 (Table 1)
GPT-5.6-sol0.0740.0680.0940.0740.078 (Table 1)
SDG (trained on SDG-30K)0.1460.1160.1620.1190.137 (Table 1)

On every generator, VIEScore2 ranks second to fourth among the 12 methods of Table 1, above all closed models and all prior trained evaluators except SDG, and GRPO helps on all four. SDG, whose detector is trained on SDG-30K, is first. Compared with SANA-1.5, the earliest of the four, VIEScore2's grid IoU on the three later generators is 0.015 lower on average. The other methods change by −0.032 to +0.007 over the same split, so this difference is not specific to VIEScore2.

FLUX.2 [klein] and Z-Image. The samples at the top come from the two models the reviewer suggested. They have no defect labels, but VIEScore2's score still orders them sensibly (table at the top), and it localizes visible defects such as misspelled letters.

Unseen sources. A new leave-one-dataset-out experiment (LoRA SFT on a 15% subsample; full table in our response to Reviewer AAhw) shows that VIEScore2 transfers partially to a source it never sees: on ImagenWorld editing, grid IoU is 0.167 without any ImagenWorld training data, against 0.296 with it.

Figure 3 annotations (Q3).

We thank the reviewer for pointing out these ambiguities. Both examples use the datasets' original annotations, converted to our evaluation format. The original annotation covers the entire banner. Figure 3 displays its 16×16 grid conversion, where cells with less than 50% overlap are excluded. Against the original polygon, VIEScore2 still achieves higher pixel IoU than GPT-5.6-sol (0.54 vs. 0.40), whose prediction extends beyond the annotated region. SDG-30K. The annotation identifies a subtle prompt mismatch: the requested "symmetrical face" conflicts with the asymmetric gold inlays. While difficult to discern, VIEScore2 localizes the annotated region (grid IoU 0.61). We will include the original annotations alongside their grid conversions in the revised figure and clarify the ambiguity of these examples.

We hope these experiments and results have adequately addressed reviewer’s concerns, and we sincerely appreciate the reviewer’s consideration of these improvements in their final evaluation.

Reviewer_dwKaOpenReview ↗

Response to Authors

Thank you for the detailed responses and additional experiments, which have addressed my main concerns regarding reward modeling and generalization to recent generators.

Regarding Q3, I would like to further discuss how defect identification aligns with human preferences. In my practical use of image generation models, the subtle asymmetry in the SDG-30K porcelain figure in Figure 3 seems acceptable: when using “symmetrical face,” I might intend broadly similar patterns on both sides rather than strict mirror symmetry. In contrast, the unclear text in the SynthScars example seems less acceptable, and I would prefer the localization to cover the entire defective text region. These examples suggest a possible distinction between matching dataset annotations and identifying defects that users actually care about. While these are only a few examples and reflect my personal preferences, they leave me uncertain about this aspect of VIEScore2. Could the authors clarify what evidence supports its alignment with human judgments of defect presence, extent, and importance, beyond overall score correlation and agreement with existing spatial annotations?

We thank the reviewer for the follow-up. We agree that agreement with benchmark annotations does not necessarily imply the detection of perceptually meaningful defects. To address this, we conducted a user study evaluating defect validity, perceived importance, and localization quality.

User Study. We sampled 30 images: 24 from four held-out benchmarks (SynthScars, SDG-30K, HAD, and AbHuman; 6 each) and 6 from in-domain test sets (RichHF and ImagenWorld; 3 each). 8 raters independently evaluated VIEScore2's predictions and benchmark annotations, presented in randomized, unlabeled order.

VIEScore2AnnotationKrippendorff's α
Q1: Defect presence63%60%0.37
Q2: Perceived importance62%63%0.02
Q3: Full coverage / Over-localization38% / 34%43% / 24%0.17
Q4: Missed defects20%24%0.20
Q5: Preference (decided pairs)53%47%0.14

At least seven of eight raters agree on defect presence for 30 of 60 markings (12 unanimously).

Takeaway. VIEScore2 matches benchmark annotations on defect presence and importance and misses fewer defects, at the cost of slightly broader regions; raters agree on whether a defect exists but not on how much it matters (α ≈ 0), so importance is inherently subjective. Consistently, on RichHF's graded human heatmaps, VIEScore2's recall rises from 0.24 to 0.66 with defect intensity (GPT-5.6-sol: 0.39 to 0.50).

Regarding the two examples:

  • Porcelain figure: VIEScore2 identifies deviations from the prompt's explicit "symmetrical face", consistent with the SDG-30K annotation. Whether such asymmetry is acceptable is subjective, as the low importance agreement above shows.
  • SynthScars: VIEScore2 covers the full text width but only its lower part, because the text spans two grid rows and the annotation converted to 16×16 keeps only the bottom row. This under-coverage is an artifact of the grid label rather than typical behavior: in the user study, VIEScore2 leaves defects partially covered or missed less often than the benchmark annotations (28% vs. 33%).
Reviewer_AAhwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VIEScore2, a unified model for image quality scoring and defect localization. It uses a 16×16 sparse grid to represent defects and trains a VLM with SFT followed by GRPO. Experiments show improved scoring correlation and competitive defect localization across multiple benchmarks.

Critical strengths
  1. The paper combines image quality scoring and defect localization in one model, and supports both image generation and editing tasks, including those with multiple reference images. This makes the framework useful for a wider range of tasks.

  2. The experimental results are quite good. VIEScore2 achieves the best Grid IoU on 3 out of 6 benchmarks (Table 1) and outperforms the general-purpose VLM baselines in overall scoring correlation (Table 2).

  3. The ablation studies are fairly thorough. The authors compare SFT and GRPO, different reward components, and grid resolutions. The results show that GRPO improves localization more than simply adding another SFT epoch (Tables 18–19).

Critical weaknesses
  1. Sparse Grid (Tables 22–23): Why use a 16×16 grid instead of masks or bounding boxes? It loses spatial details, and bounding boxes actually perform better in Table 22.

  2. GRPO Reward (Section 3.4):The Dice reward uses the union of artifact and misalignment regions, so predicting the wrong defect category is not penalized.

  3. False Positives (Table 21):GRPO improves localization but also increases false positives on clean images from 63/115 to 85/115.

  4. Comparison with ImageDoctor (Tables 1, 9): ImageDoctor already does scoring, localization, and explanation with RL. It is unclear how much of the improvement comes from the proposed method rather than different training data or backbones.

  5. Dataset Consistency (Table 4): The datasets have different types of defect annotations. Simply converting them into the same format does not guarantee that they use consistent definitions of defects.

  6. Explanations (Section 3.5, Figure 7):The explanations mostly describe where the predicted defects are, without explaining what is actually wrong with those regions.

Questions for authors
  1. In Table 22, bounding boxes perform better than sparse grids. Why did the authors choose sparse grids? Have they tried bounding boxes with GRPO?

  2. Why does the GRPO reward combine artifact and misalignment regions instead of treating them separately? Would this make it harder to distinguish the two types of defects?

  3. In Table 21, false positives on clean PAL4VST images increase from 63 to 85 after GRPO. Why does this happen? Is there a trade-off between localization accuracy and false positives?

  4. The datasets use different defect annotations (Table 4). Are the definitions of defects consistent across these datasets? How is this handled during training?

3 comments
AuthorsOpenReview ↗

Response to AAhw (1/2)

We are grateful to the reviewer for the careful reading and the precise questions. Below, we address each point in turn.

Grids vs. bounding boxes (W1, Q1).

Masks have no compact text form. Cell indices serve both as the SFT target and as an exactly scorable GRPO reward, and they can represent disconnected regions. At N = 16, the grid keeps a pixel-IoU ceiling of 0.615 at 120 tokens on average (Appendix Table 23). Table 22 shows that boxes are ahead of cells after SFT alone (grid IoU 0.280 vs. 0.246). As the reviewer suggested, we also trained both formats with GRPO under the same 9,075-example corpus and settings, rewarding both against the same ground-truth grid (IoU for boxes, Dice for cells).

New experiment: Grid IoU, primary suiteCellsBoxesΔ (Cells − Boxes)
SFT0.2540.275−0.021
SFT + GRPO, trial 10.2770.271+0.006
SFT + GRPO, trial 20.2840.282+0.002

Boxes are again ahead after SFT, but GRPO improves only the cell format, so the two end level. We will add this experiment.

Union Dice reward (W2, Q2).

Only RichHF labels the two categories separately, so the union gives every source the same reward. In the paper, GRPO with the union reward already improves both RichHF categories over SFT: artifact F1 rises from 0.488 to 0.493, and misalignment F1 from 0.165 to 0.198 (Table 21). As a direct test, we trained a category-aware variant, whose reward on RichHF examples is the mean of the artifact and misalignment Dice, and a union-reward reference from the same re-trained SFT checkpoint.

New runsGrid IoU, primary suiteArtifact F1Misalignment F1
GRPO, union reward0.3200.4650.184
GRPO, category-aware reward0.3230.4910.198

It leaves union grid IoU essentially unchanged and improves both categories slightly, so the union reward costs little in category separation. We will report the category-aware variant as an option.

False positives on clean images (W3, Q3).

This is a tunable precision/recall trade-off, and the increase is smaller than the image count suggests. In Table 21, GRPO raises the number of flagged clean PAL4VST images from 63 to 85 of 115, while false alarms on COCO photographs stay at 0 of 100. A new analysis shows that the 22 additional images contain a median of 2 predicted cells and that total false-alarm cells fall from 1,357 to 819. PAL4VST's "clean" labels are also incomplete: 450 of its 507 "clean" test images are flagged by at least 6 of 11 other evaluators, including PAL4VST's own specialist. Finally, the reward can move the operating point: in Table 20 (development set), a precision-oriented Tversky reward (α = 0.7) lowers false alarms on clean PAL4VST images from 26 to 21 of 40, at a cost of 0.009 grid IoU (0.322 → 0.313).

Comparison with ImageDoctor (W4).

We agree that the backbone and data differ, and we will not attribute the full gap to the method. First, the backbone alone is weak: zero-shot Qwen3-VL-8B reaches 0.022 grid IoU on RichHF, against 0.284 for ImageDoctor (Table 1). Second, VIEScore2 leads on RichHF and on four datasets that neither method trains on:

Grid IoU (Table 1)RichHFAbHumanHADSynthScarsSDG-30K
VIEScore20.2990.2150.1970.2340.102
ImageDoctor0.2840.1440.1580.1690.086

Third, with the backbone and data fixed, training matters: GRPO adds 0.028 grid IoU over SFT (0.296 → 0.324), whereas one more SFT epoch adds 0.001 (Table 19).

Dataset consistency (W5, Q4).

We do not assume identical defect definitions: per-source prompts state which fields each source supervises, and the grid fixes only the output format. To test for conflicts, we ran a new leave-one-dataset-out study (LoRA SFT on a 15% stratified subsample, one uniform prompt). The table gives grid IoU when training on all sources and the change when one source is removed (bold: the test set's own source).

Evaluated onAll sources− RichHF− PAL4VST− ImagenWorld
RichHF0.184−0.126−0.020−0.009
PAL4VST0.092−0.044−0.055−0.022
ImagenWorld, T2I0.337−0.028−0.050−0.110
ImagenWorld, edit/ref.0.296−0.010+0.012−0.129

Each test set drops most when its own source is removed, although transfer is partial: ImagenWorld editing still reaches 0.167 without any ImagenWorld data. No removal improves another test set by more than 0.012. The heterogeneous labels are therefore complementary rather than conflicting. We will add the study and discuss definition differences as a limitation.

AuthorsOpenReview ↗

Response to AAhw (2/2)

Explanations (W6).

We agree. The reports state where the defects are, their category, and their severity, but not what is visually wrong. Fixed rules produce them from the predictions, so they cannot contradict the model, unlike free-form rationales. We will call them structured reports, narrow the claims, and note that a VLM conditioned on the predicted cells can add free-form descriptions.

Thank you again for the detailed and constructive review.

Thank you for the clarifications, and I maintain my original score after considering the rebuttal.

Reviewer_bMzAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces VIEScore2, a unified evaluator for image generation and editing. It takes the generated image, text prompts and optional conditioning images as input, and outputs evaluation scores and defect locations including misalignment and artifacts. For defect localization, VIEScore2 represents the image as an NxN grid and predicts the row and column coordinates of the defects. VIEScore2 is trained on a joint datasets of 38K examples across multiple image generation and editing evaluation tasks. It is trained first with supervised fine-tuning and then GRPO with cell-level Dice overlap, score accuracy, and output-format rewards. Finally, VIEScore2 uses a parameter-free parser to convert the structured predictions into readable explanations. The paper evaluates the proposed method on various defect localization and scoring tasks, and demonstrates promising results.

Critical strengths
  1. The paper proposes a unified framework for joint scoring and defect localization with optional conditioning images for various image generation and editing tasks.
  2. The proposed model is jointly trained across heterogeneous datasets. The two stage training from SFT to RL improves model performance.
  3. Experiments across multiple datasets and tasks demonstrate promising results compared to baseline image evaluators and API-based models.
Critical weaknesses
  1. The writing of the paper needs to improve with many details unclear. For example, what is "primary suite" in the abstract (L22)? Is it the held-out example described in L269-292? What does "model-free pixel-IoU" mean in Fig 4? How SFT is conducted with datasets of different types of groundtruth labels?
  2. The novelty of the proposed method seems limited. The major components of VIEScore2 are already covered in prior work. For example, ImageDoctor already uses VLM to jointly predict quality scores and defect maps, and also conducts two stage SFT + RL training. And representing segmentation with textual grid outputs is introduced by Text4Seg. The main novelty seems to be the joint training of multiple image generation and editing tasks, which is incremental.
  3. The paper highlights the overall-score SRCC of 0.601, compared with GPT-5.6 and Gemini-3-Flash in Table 2. However, the evaluation suite comes from the held-out examples of the model's training data, which weakens the comparison.
  4. The paper claims "GRPO IMPROVES LOCALIZATION BEYOND ADDITIONAL SFT". However, SFT is conducted with LoRA while GRPO is run with full-parameter, which makes it hard to directly compare the performance of the two.
  5. The usefulness of the parameter-free parser is unclear. It uses predefined rules and templates to convert the structured predictions into text, but it does not explain why the model makes certain predictions. For example, why the model assigns a particular score and why a certain region does not look right.
Questions for authors
  1. Can the authors explain what is the main novelty and contribution of the paper compared to existing works? The two of main contributions listed in the introduction, i.e., "Verifiable spatial post-training" and "Learning and evaluation across heterogeneous sources." seem trivial.
  2. How SFT is conducted with datasets of different types of ground truth labels? Do you specify what to predict in the input prompts?
3 comments
AuthorsOpenReview ↗

Response to bMzA (1/2)

We thank the reviewer for their insightful comments. After addressing the questions you raised, we conducted new experiments and made revisions. We believe these efforts have strengthened the paper. Below, we address each point in turn.

Clarification on Details (W1, Q2, W3).

W1) The primary suite (L22) is the held-out set described in L269–292. It is the one containing 1300 examples (RichHF 400, PAL4VST 300, EvalMuse 200, COCO 100, ImagenWorld 300). Model-free pixel-IoU refers to the IoU between an original pixel-level annotation and the same annotation after conversion to an N×N grid. We wanted to emphasize that the calculation is deterministic. It is the upper bound that any grid predictor can reach at resolution N (0.615 at N = 16). We will rename it "grid representation ceiling". We will clarify them in the revision.

Q2) Training with different labels: Each dataset's prompt requests only the target fields supported by its annotations (Appendix Table 4):

SourceRequested fields
ImagenWorld, COCOPQ/SC scores and a unified defect grid
RichHFPQ/SC scores and separate artifact/misalignment grids
PAL4VSTDefect grid only
EvalMuseOverall score only
All spatial annotations are mapped onto the same 16 x 16 grid. During SFT, loss is computed strictly over the requested tokens. During GRPO, reward terms lacking ground-truth supervision receive zero weight and contribute no advantage.

W3) Human Correlation: We thank the reviewer for raising this concern. We agree that the primary suite is in-distribution of the training sources. We will explicitly clarify it in our paper.

This is the standard way to test trained evaluators (e.g., RAHF, ImageDoctor, FGA-BLIP2). All these methods have held-out training data as test data. To further assess generalization, we conducted additional evaluations in the paper:

  • Scoring: We have 0.773 SRCC on a disjoint 2,000-example EvalMuse subset.
  • Localization: On four unseen datasets during training, grid IoU reaches 0.215 on AbHuman (3rd/12), 0.197 on HAD (2nd), 0.234 on SynthScars (1st), and 0.102 on SDG-30K (4th in IoU, 2nd in F1).

Novelty Concerns (W2, Q1).

We acknowledge that joint scoring and localization, SFT followed by RL pipeline, and text-based segmentation have precedents. Our contribution is not these components individually, but their integration and validation for broader image evaluation tasks:

  1. Conditional evaluation. Unlike ImageDoctor, which evaluates prompt-image pairs, VIEScore2 also supports editing and multi-reference generation using source and reference images.
  2. Verifiable localization. Our sparse-cell representation enables exact cell-level Dice rewards. Under identical GRPO training, grid IoU improves by +0.024 with cells, compared with −0.004 for boxes (R2, W1).
  3. Heterogeneous supervision. A new leave-one-dataset-out study (LoRA SFT on a 15% stratified subsample) shows that each dataset has its own complementary contributions. Removing any one of them worsens the results in most cases.
Grid IoU onAll sources− RichHF− PAL4VST− ImagenWorld
RichHF0.184−0.126−0.020−0.009
PAL4VST0.092−0.044−0.055−0.022
ImagenWorld, T2I0.337−0.028−0.050−0.110
ImagenWorld, edit/ref.0.296−0.010+0.012−0.129

These design choices also yield practical gains. On RichHF, VIEScore2 reaches 0.299 grid IoU versus 0.156 for GPT-5.6-sol under matched inputs (Table 1). Beyond evaluation, new experiments in our response to Reviewer R3 show that VIEScore2 is a useful reward: used to fine-tune FLUX.2-klein with DiffusionNFT, it improves all four independent metrics over the base model and matches or slightly exceeds PickScore under the same budget. We will clarify these contributions in the revision.

SFT vs. GRPO parameterization (W4).

We agree that the original comparison changed the parameterization. We ran GRPO with LoRA (rank 32, as in SFT):

Primary suiteGrid IoUF1Δ IoU vs. its SFT init
SFT (LoRA), epoch 40.2960.471—
+ 1 more SFT epoch0.2970.472+0.001
+ GRPO, full-parameter0.3240.506+0.028
+ GRPO, LoRA (new)†0.3170.495+0.043
† indicates that it is trained from a re-trained SFT checkpoint. The grid IoU on the re-trained checkpoint is 0.274. The full-parameter GRPO re-run from that checkpoint reaches 0.320 (+0.046).

With the same parameterization as SFT, GRPO gains 0.043 grid IoU and matches full-parameter GRPO from the same checkpoint (0.317 vs. 0.320), whereas an extra SFT epoch gives 0.001. The gain therefore comes from GRPO, not from full-parameter training. We will add the LoRA row.

AuthorsOpenReview ↗

Response to bMzA (2/2)

Parameter-free parser (W5).

We clarify that the parser does not aim to explain the model's underlying reasoning. Instead, it provides faithful, human-readable reports in which every statement is directly traceable to a predicted cell group, defect category, or score. This ensures consistency between textual and structured outputs, allowing users to verify each claim against the predictions and downstream systems to reliably act on the reported defects. It also avoids unsupported claims that may arise from free-form rationales.

We hope that the revisions have effectively addressed the concerns raised and further improved the manuscript, and we would greatly appreciate the reviewer’s consideration of these changes in their final evaluation.

Dear Reviewer bMzA,

Thank you again for your thoughtful and constructive feedback on our paper. In response to your comments, we have conducted additional experiments and provided further clarifications to address your concerns.

As the author–reviewer discussion period is drawing to a close, we would greatly appreciate it if you could let us know whether our responses have satisfactorily addressed your concerns or if there are any remaining questions.

We sincerely appreciate your time, effort, and valuable insights throughout the review process.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank the reviewers for their valuable feedback. In response to these suggestions, we have added new experiments and analysis into the revised version. The main text is kept unchanged. The extended results have been added as new rows to existing tables, and new experiments that could not be accommodated in the existing appendix have been collected in a new Appendix H. The specific changes are summarized below:

Training and Reward Design

  • We implemented GRPO with LoRA (the same parameterization as SFT) and compared it against full-parameter GRPO; both were initialized from the same SFT checkpoint (Reviewer bMzA; Appendix H.1, Table 24). The results demonstrate that the localization gain stems from GRPO itself, rather than from full-parameter training.
  • We introduced a category-aware Dice reward (Reviewer AAhw; Appendix H.1, Table 24). The union reward maintains overall performance while resulting in only a negligible loss in category separation.
  • Under identical settings, we compared GRPO training using bounding boxes versus sparse cells (Reviewer AAhw; Appendix H.2, Table 25). As GRPO is specifically optimized for the cell format, both formats ultimately achieve comparable performance.

Generalization

  • Regarding the training sources, we conducted a leave-one-dataset-out study (Reviewers bMzA and AAhw; Appendix H.3, Table 26). The results indicate that the heterogeneous sources provide complementary, rather than conflicting, supervision.
  • We performed a per-generator breakdown of SDG-30K (Reviewer dwKa; Table 17, Appendix E.1).
  • We evaluated the model on new samples from FLUX.2 [klein] and Z-Image, neither of which appears in our training data (Reviewer dwKa; Appendix H.4, Table 27, Figure 11).

Downstream Use

  • DiffusionNFT fine-tuning of FLUX.2-klein with VIEScore2 or PickScore as the reward (Reviewer dwKa; Appendix H.5, Table 28).

Clarifications

  • A further analysis of false alarms on clean PAL4VST images (Reviewer AAhw; Appendix E.1).
  • The original annotation of the SynthScars example in Figure 3, along with its grid conversion (Reviewer dwKa; Appendix H.6, Figure 12).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.