acceptodds
Under review as a conference paper at ICLR 2027

ReflowText: Beyond Local Text Replacement for Coherent Visual Content Revision

Abstract

Image editing models have become effective at replacing short words and phrases, but real updates often involve long text, source-to-target length changes, and multiple jointly edited regions. Beyond accurate glyphs, models must adapt line breaks, text scale, occupied regions, and surrounding layout while preserving style and non-edited content. Existing benchmarks do not systematically isolate these challenges. We introduce , an English-to-English, replacement-only benchmark with 2,002 full images and 9,197 regions across four categories and 15 subcategories. It controls three coupled difficulty axes—long-form content, multi-region replacement, and length mismatch—and combines OCR metrics with a fine-tued VLM judge, , to evaluate text fidelity, font style, layout and reflow, and background preservation. ReflowJudge reaches 79.1% agreement with human annotations. We further propose , a plug-and-play layout-first strategy in which a fine-tuned VLM predicts line-level layouts, a deterministic renderer creates spatially aligned glyph conditions, and the image editor restores visual appearance. Across two open-weight backbones, GlyphPlan consistently improves aggregate quality, especially font style and layout. Together, ReflowText and GlyphPlan advance reliable long-form, multi-region visual text editing with coherent layout adaptation.

Primary area:
datasets and benchmarks
Keywords:
text replacement, scene text, image editing, long-form text, layout reflow, glyph conditioning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Us3UOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces ReflowText, an English-to-English visual text editing benchmark focusing on long-form content, multi-region replacement, and source-to-target length mismatches. It also presents ReflowJudge, a VLM-based evaluator that complements OCR-based metrics by assessing font-style consistency, layout and reflow quality, and background preservation. Furthermore, the authors propose GlyphPlan, a VLM-based layout planner that predicts line-level text arrangements and provides spatially aligned glyph conditions to guide the image editing process. Experiments on two open-source editing backbones demonstrate improvements in editing quality, particularly in font-style consistency and layout.

Critical strengths
  1. The paper focuses on a relatively underexplored setting in visual text editing, particularly long-form text replacement, length mismatches, and layout adaptation. The proposed ReflowJudge also provides a task-specific evaluation framework based on an open-weight VLM, which could be useful for future research and benchmarking.

  2. The paper explores a layout-first approach using a VLM to predict line-level text placement before image editing. This provides an explicit way to separate layout planning from visual rendering, and the experiments show some improvements in font-style consistency and layout quality.

Critical weaknesses
  1. Although ReflowText covers several important challenges in visual text editing, the paper mainly presents overall performance comparisons and a few qualitative examples. I would like to see more detailed analyses of how different models perform under varying text lengths, length mismatches, and numbers of edited regions. Combined with representative visualizations, these analyses could help identify the main failure modes and limitations of current editing models, making the benchmark more informative.

  2. Previous works, such as TextDiffuser-2 and WeEdit, have explored language-model-based layout planning for visual text generation or editing. While GlyphPlan offers a trainable line-level planning approach, I'd like to understand how much benefit comes from the planner's specialized training. For example, how would a strong yet affordable zero-shot VLM (e.g., GPT-6 Luna) perform as the layout planner in the same editing pipeline? Such a comparison would help clarify the practical advantages of GlyphPlan. Additional details about the planner's training data would also be useful.

Questions for authors
  1. Could the authors provide more detailed results across different text lengths, length mismatches, and numbers of edited regions? It would be helpful to understand where current models tend to fail and which challenges remain the most difficult. Representative failure cases would also be useful.

  2. Have the authors considered using a strong zero-shot but affordable VLM (e.g., GPT-6 Luna) to generate the line-level layouts instead of training a dedicated planner? Such a comparison could help clarify the benefits of the proposed training strategy.

  3. GlyphPlan uses a standard font to render glyph conditions without explicitly encoding the original typography. However, Table 5 shows noticeable improvements in font-style consistency. Could the authors further explain this observation? In particular, I would be interested in seeing how GlyphPlan performs when the original font style differs substantially from the standard glyph font, such as highly decorative or handwritten text. Some qualitative examples or a breakdown of performance in such cases would help demonstrate whether spatially aligned glyph conditions are sufficient for preserving diverse font styles.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Us3U

Thank you for the constructive suggestions and for highlighting the font-style improvement. We have added stratified results, controlled planner comparisons, and further analysis of font-style preservation.

Q1. Results stratified by difficulty.

Appendix B.3 adds text-load and replacement-length distributions. Appendix G further breaks down the results of the 14 models in the main paper’s Table 3 by text load, replacement-length mismatch, number of edited regions, and domain. The table below reports equally weighted mean Final scores across these models on a 0–100 scale, in the same order as the groups. Text load sums target characters across all edited regions in an image.

AnalysisGroups, in orderMean Final ↑
Target characters per image≤128 / 129–512 / >51267.28 / 55.94 / 41.00
Target-source word-count ratio per image<1 / =1 / >155.29 / 65.87 / 58.74
Edited regions per image1–2 / 3–4 / 5–6 / 7–863.54 / 58.38 / 56.50 / 54.65
Primary domainDigital interfaces / Graphic design / Natural scenes / Documents66.37 / 57.61 / 56.45 / 51.53

Full per-model results appear in Tables 16–18, with visual summaries in Figures 12–15.

All 14 editors score lower in the highest versus lowest text-load and region-count groups, highlighting the difficulty of accurately rendering longer text and coordinating multiple replacements. Both contraction and expansion also score below equal total word count, as length mismatch can disrupt alignment and complicate reflow. Across domains, documents are the hardest; newspapers, academic papers, and reports score 37.52, 45.75, and 47.31, respectively, compared with 72.12 for presentation slides (Fig. 14), highlighting the challenges of text-dense content. Figure 16 presents typical failure cases: long-text replacement produces garbled text and incorrect characters, length mismatch causes layout errors, background elements are lost, and decorative or handwritten text suffers from font-style inconsistencies.

Q2. Can a strong zero-shot VLM serve as the planner?

We compare GlyphPlan with six alternative planners: three closed-source API models, including GPT-6 Luna, and three open-weight VLMs without task-specific fine-tuning. The renderer, with-glyph SFT editor, editing prompt, and sampling settings remain fixed. Only the planner changes, and all 2,002 images remain in evaluation on SenseNova-U1.5 and Qwen-Image-Edit-2511. Final scores are summarized below; complete results for all six evaluation metrics and Final are provided in Appendix H.1, Tables 19–20:

PlannerSenseNova Final ↑Qwen Final ↑
GlyphPlan (trained)66.2853.19
GPT-6.1 Sol65.4051.82
Gemini 3.8 Flash64.8749.33
gemma-4-31B-it64.6249.12
Qwen3.6-27B (base)63.6247.54
Qwen3.5-9B63.1844.72
GPT-6 Luna61.7447.76

GlyphPlan improves Final over GPT-6 Luna by 4.54/5.43 points and over its unadapted Qwen3.6 planner base by 2.66/5.65 (SenseNova/Qwen). GPT-6.1 Sol is competitive on individual metrics, but GlyphPlan has the highest Final on both editors. API planners require recurring service charges; GlyphPlan can be deployed locally.

Q3. Why can standard-font glyphs improve handwriting and decorative styles?

Table 5 shows improved Style on both editors under matched SFT settings:

EditorStyle: SFT without glyphsStyle: SFT + GlyphPlan
SenseNova-U1.562.2373.06
Qwen-Image-Edit-251154.8760.02

We appreciate your suggestion to examine challenging font styles and have followed it by evaluating examples from Natural Scene – Handwritten Casual, comparing ordinary zero-shot, no-glyph SFT, and SFT + GlyphPlan editing (Fig. 17, Appendix I.1). The visualizations show more natural handwritten and decorative lettering with GlyphPlan, with better fidelity to the source font style, color, and character proportions than the baselines.

GlyphPlan decouples text-layout planning from image editing, allowing the editor to devote more expressive capacity to font-style preservation and focus more on the original lettering in the source image. Our attention experiment supports this explanation: measuring attention from queries in the generated editing region to source-image keys, both editors allocate a larger share to source-text regions on average across all 2,002 benchmark images (Table 22, Appendix I.2):

EditorSFT without glyphSFT + GlyphPlanGain (percentage points)
SenseNova-U1.537.17%38.85%+1.68
Qwen-Image-Edit-251155.86%57.31%+1.45

The attention maps in Figure 18 clearly show increased attention concentrated on the original text regions, alongside more natural handwriting in the edited output. This suggests that greater attention to the source lettering helps the editor retain its font-style features.

Thank you for the thorough rebuttal. Your new experiments in Tables 16–20 address Q1 and Q2 well. However, my concern about Q3 remains:

  • A ~1.5% rise in source-text attention (Table 22) is quite small to explain the large Style gain.
  • In ReflowJudge (Appendix D.3), the Style metric also checks basic character legibility (glyph integrity), so much of the gain may simply come from cleaner text rendering rather than better font-style transfer.
  • In Figure 17 ((b), (c), and (f)), SFT + GlyphPlan still turns handwritten and chalkboard text into standard printed fonts.

To clarify this without running new experiments, could you report the Style scores from Table 5 (SFT no glyph vs. SFT + GlyphPlan) specifically on stylized subcategories like Handwritten Casual and Movie Poster from your existing logs? Alternatively, if you have other effective ways to demonstrate this point, those would also be very welcome.

AuthorsOpenReview ↗

Further Evidence and Analysis of Font-Style Preservation

Thank you for the prompt follow-up and for recognizing that the new experiments address Q1 and Q2. We apologize that our previous response did not fully resolve your concern about Q3. We respond below.

1. Evidence that GlyphPlan improves font-style preservation.

We agree that glyph integrity contributes to ReflowJudge's Style criterion. However, the large Style gains in Table 5 accompany only small changes in foreground text accuracy. Each entry below shows SFT without glyphs → SFT + GlyphPlan:

EditorFG-NED ↓FG-Recall ↑Style ↑
SenseNova-U1.519.70 → 18.5757.91 → 57.1062.23 → 73.06
Qwen-Image-Edit28.97 → 28.4637.53 → 38.2954.87 → 60.02

In particular, SenseNova gains 10.83 Style points while FG-Recall slightly decreases; Qwen gains 5.15 Style points with only a 0.76-point increase in FG-Recall. These results indicate that glyph integrity is not the main driver of the Style gains and does not alter our conclusion that GlyphPlan improves font-style preservation.

Following your suggestion, we also extracted Style scores for Handwritten Casual and Movie Poster from our evaluation logs:

SubcategoryEditorSFT without glyphsSFT + GlyphPlan
Handwritten CasualSenseNova-U1.549.3559.16
Handwritten CasualQwen-Image-Edit53.8263.36
Movie PosterSenseNova-U1.552.3360.51
Movie PosterQwen-Image-Edit50.1956.91

The gains persist on both stylized subcategories and both editors. We agree that the font styles in Figure 17 still retain some AI-generated characteristics and are not yet perfect. Nevertheless, GlyphPlan preserves the original font style substantially better than no-glyph SFT. For example, the word “Thanks” in Figure 17(b) looks noticeably closer to natural handwriting with GlyphPlan than with no-glyph SFT. We attribute the remaining tendency toward standard fonts to the editors' learned appearance priors and the limited representation of difficult fonts in our training data. GlyphPlan alleviates this limitation without fully resolving it. Further addressing it calls for dedicated research, such as handwriting generation[1] and artistic typography[2], together with substantial training data tailored to these styles.

2. Feature-level analysis toward interpreting the mechanism behind Style gains.

We agree that the 1.45–1.68 percentage-point attention increase alone is insufficient to establish the proposed mechanism behind the Style gains. To explore the mechanism from another perspective, we compare source–output text-region similarity using DINOv2 and diffusion features. Detailed experimental settings and full results are provided in Appendix I.3, Tables 23–24.

EditorFeatureWithout glyphsWith GlyphPlanRelative gain95% CI of absolute gainHolm-adjusted p
SenseNova-U1.5DINOv20.790960.80574+1.87%[0.01137, 0.01828]1.87e-14
SenseNova-U1.5Diffusion0.804040.81956+1.93%[0.01246, 0.01860]1.54e-20
Qwen-Image-EditDINOv20.787040.81116+3.06%[0.01988, 0.02859]4.77e-24
Qwen-Image-EditDiffusion0.794000.81105+2.15%[0.01282, 0.02138]7.43e-13

All four improvements remain significant after Holm correction. A fixed-region control over all 2,002 images also shows significant gains for both editors and both feature types.

The layer-wise results are consistent (Table 24). Entries show SFT without glyphs → SFT + GlyphPlan, with relative gains in parentheses:

Transformer layerSenseNova-U1.5Qwen-Image-Edit
60.83622 → 0.85083 (+1.75%)0.83066 → 0.84436 (+1.65%)
180.82854 → 0.84343 (+1.80%)0.81998 → 0.83488 (+1.82%)
300.83789 → 0.85165 (+1.64%)0.83114 → 0.84405 (+1.55%)
410.80404 → 0.81956 (+1.93%)0.79400 → 0.81105 (+2.15%)
530.79621 → 0.80775 (+1.45%)0.79524 → 0.80755 (+1.55%)

Both editors show positive mean gains across all five reported transformer layers. Together with the attention maps, these results support better source-appearance preservation across feature types and layers.

Thank you for your constructive feedback and guidance. We hope that our response and additional evidence address your concerns.

[1] Nikolaidou, Konstantina, et al. "Diffusionpen: Towards controlling the style of handwritten text generation." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

[2] Tanveer, Maham, et al. "Ds-fusion: Artistic typography via discriminated and stylized diffusion." 2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2023.

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

This paper looks at text replacement in images when the new text is much longer or shorter than the original and several regions need to change together. In this setting the model has to rethink line breaks, text size and placement instead of simply swapping words. The authors make two contributions. The first is ReflowText, a replacement-only English benchmark built from real images across scenes, posters, interfaces and documents. Target texts are generated to be shorter than, similar to or longer than their sources. Evaluation combines OCR-based text accuracy with a fine-tuned VLM judge that checks style, layout and background preservation, and the judge is validated against human labels. The second is GlyphPlan, which first predicts a line-level layout with a VLM, renders it as a glyph image, and then uses a fine-tuned editor to produce the final result. Experiments benchmark a range of open and proprietary editors and compare different glyph conditioning strategies on two open backbones. The results suggest that a well-aligned layout plan helps, while a poorly aligned glyph condition can hurt.

Critical strengths
  1. The paper identifies a real gap. Existing text editing benchmarks mostly test short or length-matched swaps. ReflowText makes length mismatch and multi-region editing the default: 97.9% of images need at least two edits, and the images are balanced over four domains. The evaluation of 14 recent editors shows that strong systems still fail on overflow, style drift and background changes, which earlier benchmarks were not built to expose. This is the main reason I see value in the paper.

  2. The controlled glyph ablation is informative. With data and compute matched, naive glyphs placed in the source box reduce Final by 9.31 on Qwen and 3.09 on SenseNova, while planned glyphs help. The message is clear: glyph conditioning only helps when its layout agrees with the target. This is useful for anyone designing glyph-based editors.

Critical weaknesses
  1. No stratified results. The paper says ReflowText controls three difficulty axes, but none of the results are broken down by length policy, target length, number of regions, or domain. For a benchmark, this analysis is the main thing readers need, and without it the axes are only described, not tested. The median target is also fairly short, so the long-form claim relies mostly on the tail of the distribution.

  2. Loophole in the OCR metrics. FG-NED and FG-Recall search the whole OCR transcript, so they do not check whether a target lands in the right region. Some targets are already contained in their source after normalization, for example "tree branch" to "branch" and "Buy." to "Buy". An editor that returns the input unchanged gets credit on these pairs. Text fidelity carries most of the weight in Final, so this can affect the rankings.

  3. Background judgments may reflect the judge more than the editors. On Background, ReflowJudge rejects about half of the outputs that humans accept, and every system scores low on this dimension. It is hard to tell whether editors really damage the background this often or whether the judge is too strict.

  4. The human validation is small and does not match how the metric is used. Only 50 images are annotated, and Style and Layout seem to be labeled once per image, while the benchmark scores them per block. Inter-annotator agreement before adjudication is not reported.

  5. The GlyphPlan claims go beyond the evidence. In a benchmark paper the baseline does not need to be state of the art, but the conclusions should match the results. The gain is about 2 points from a single run without confidence intervals. On SenseNova, the final model is only about 1 point above the untuned editor, and FG-Recall drops slightly. Zero-shot use hurts both backbones, so calling the method plug-and-play is not accurate.

  6. Training data source is not described. The SFT pairs need target images, but the paper does not say where they come from. If a proprietary editor produced them, the SFT is partly distillation. Overlap with the benchmark's source datasets should also be checked.

  7. The planner is not evaluated on its own. There is no line break accuracy, box IoU, or oracle layout upper bound, and no report of the extra inference cost. Axis-aligned line boxes also cannot represent curved or perspective text, which is common in natural scenes.

Questions for authors
  1. Could you report the main results split by length policy, number of regions and domain? This is the analysis I most want to see, since it is what the benchmark is designed to support.
  2. What does a copy-input baseline score, which returns the source image unchanged? How many pairs have a normalized target that already appears in the source? If this baseline scores well on the text metrics, the weighting of Final may need to change.
  3. In the human study, are Style and Layout labeled per block or once per image? What was the agreement between the two annotators before adjudication?
4 comments
AuthorsOpenReview ↗

Response to Reviewer MGEx

Thank you for recognizing the benchmark's value and for your valuable suggestions. In response, we have added several important experiments and clarified the points you raised.

Q1. Results stratified by difficulty.

Appendix G analyzes the 14 models in the main paper’s Table 3 by text load, length mismatch, region count, and domain. Scores below are equally weighted mean Final (0–100), listed in group order. Text load totals target characters per image.

AnalysisGroups, in orderMean Final ↑
Target characters per image≤128 / 129–512 / >51267.28 / 55.94 / 41.00
Target-source word-count ratio per image<1 / =1 / >155.29 / 65.87 / 58.74
Edited regions per image1–2 / 3–4 / 5–6 / 7–863.54 / 58.38 / 56.50 / 54.65
Primary domainDigital interfaces / Graphic design / Natural scenes / Documents66.37 / 57.61 / 56.45 / 51.53

Full results and visualizations are provided in Appendix G, Tables 16–18 and Figures 12–15.

All 14 models score lower in the highest versus lowest text-load and region-count groups, showing the difficulty of longer text and multiple replacements. Contraction and expansion underperform equal word counts, as mismatch can disrupt layout. Documents are hardest, highlighting dense-text challenges. Figure 16 presents representative failure cases: garbled text, layout errors, background loss, and font-style changes.

Q2. Copy-input and source–target overlap.

We test the shortcut by setting each predicted string equal to its annotated source, across all 2,002 images and 9,197 replacement pairs. Under this paired annotation-text control, normalized-target recovery (FG-Recall) is only 5.80%. The resulting foreground contribution is 12.89/60, well below 13 of the 14 editors. Selected results are shown below; full results are provided in Appendix B.6, Table 8:

MethodFG-NED ↓FG-Recall ↑Foreground Score ↑ (/60)
GPT-Image-215.8064.2344.53
SenseNova-U1.519.4257.2441.34
Qwen-Image-Edit-251131.7834.2130.73
LongCat-Image38.3929.9627.47
BAGEL67.655.2511.28
Copy-input text control62.855.8012.89

This control measures overlap within each annotated source–target pair; it does not run full-image OCR or report the six-metric Final. It shows that returning the source text verbatim receives little foreground credit.

Q3. Human annotation granularity and agreement.

To clarify, the human study evaluates 700 edited images, not just 50: 50 source images × outputs from 14 models. Each output is assessed on three dimensions, yielding 2,100 annotation units, each independently labeled by two annotators. For agreement, the VLM's block-level Style/Layout judgments are converted to image-level decisions using the same human rule: one failing block makes the image fail. Table 12 reports agreement before adjudication:

DimensionHuman–human agreement
Style84.4%
Layout83.9%
Background Preservation87.9%
Overall85.4%

Even human agreement is imperfect for these visual criteria. ReflowJudge's 79.1% agreement with adjudicated labels is 6.3 points below this reference, supporting its effectiveness as an automated evaluator (Appendix F.1).

Q4. Training-data provenance.

Appendix C.2 clarifies that SFT source images come from entirely different sources than the benchmark. The 91,321 pairs combine synthetic text overlays and GPT-Image-2-generated edits, with repeated filtering and quality checks.

Q5. Background strictness.

Thank you for pointing this out. We agree Background is overly strict, though its 78.9% human agreement supports its use. We suspect strict GPT-5.5 teacher data contributes (recall: 66%). We are correcting the data and refining the judger; preliminary results show recall rising from 51% to 64%. Meanwhile, our combined variant also achieves 78.9% agreement, with 77% recall versus split's 51%, and can serve as an interim Background evaluator.

Q6. Scope of the GlyphPlan claim.

We apologize for the imprecise “plug-and-play.” GlyphPlan is designed to work across editor backbones, rather than to be training-free. Table 5 shows consistent gains over matched no-glyph SFT:

BackboneΔ FinalΔ Style
Qwen+1.89+5.15
SenseNova-U1.5+1.98+10.83

The main gain is in Style, as reviewer Us3U also noted. Appendix I provides visual examples and attention analysis supporting the mechanism.

Q7. Planner evaluation and curved text.

We assess line-break and layout quality through the VLM-based Layout score on final edited images. We do not evaluate the planner separately using IoU or exact-break accuracy: there is no unique correct line break or layout, and multiple arrangements can be valid. Perspective and curvature, like font color and style, are handled by the source-conditioned editing model rather than GlyphPlan.

Dear Reviewer MGEx,

Thank you for recognizing the benchmark's value and for your constructive feedback. We summarize our main responses below:

  • Stratified results (Q1). We report results for all 14 main-table models by text load, length mismatch, region count, and domain. The results show greater difficulty with longer text, more regions, and length mismatch, with representative failure cases (Appendix G, Tables 16–18 and Figures 12–16).
  • Copy-input control (Q2). Setting each predicted string to its annotated source yields only 5.80% FG-Recall and a foreground score of 12.89/60, below 13 of the 14 editors. These results show that simply copying the annotated source text earns limited foreground credit (Appendix B.6, Table 8).
  • Human evaluation (Q3). Our study covers 700 edited outputs (50 sources × 14 models) and 2,100 output–dimension units, each labeled by two annotators. Human–human agreement is 85.4%, compared with ReflowJudge's 79.1% agreement with adjudicated labels (Appendix F.1, Table 12).
  • Other weaknesses (Q4–Q7). We also clarify training-data provenance, Background evaluation strictness, the scope of GlyphPlan’s claims, and planner evaluation and curved-text handling.

As the discussion period ends in two days, we would greatly appreciate it if you could review these updates and let us know whether they address your concerns. If any point remains unclear, we would be glad to provide further clarification. Thank you again for helping us improve the paper.

Best regards,
The Authors

Thanks for the authors' detailed response, which solved most of my concerns. One remaining point: the copy-input control in Appendix B.6 does not run full-image OCR, so I encourage the authors to also report a true copy-input baseline (scoring the source image as the output) and to temper the claim that GlyphPlan improves text accuracy. I encourage the authors to add these in the revised version. I will remain my positive score.

Thank you for your positive assessment of our paper and for your valuable suggestions. We have conducted the true copy-input experiment you recommended and will include these results in the revised Appendix B.6, Table 8.

We treat each unchanged source image as the output and run the same full-image OCR and ReflowJudge evaluation on all 2,002 benchmark images:

FG-NED ↓FG-Recall ↑BG-Text F1 ↑Style ↑Layout ↑Background ↑Final ↑
65.445.1195.4099.9699.97100.0051.55

Copy-input scores 51.55 Final, ranking 11th among 15 methods (the 14 main-table editors plus this baseline). This is expected: the unchanged image earns near-perfect Style, Layout, and background-related scores. However, its foreground score is only 11.90/60, below 13 of the 14 editors, showing that source–target text overlap does not give copy-input a high text-replacement score.

We have also revised the relevant wording in the manuscript to address your suggestion. Thank you again for helping us improve the paper.

Reviewer_xfwiOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ReflowText, an English text-replacement benchmark containing 2,002 images and 9,197 regions, targeting layout adaptation under long-form text, length changes, and multi-region editing. It combines OCR metrics with a task-specific evaluator, ReflowJudge, to assess text accuracy, font style, layout, and background preservation. The authors also propose GlyphPlan, which uses a fine-tuned vision-language model to plan line-level layouts and provides glyph conditions to adapted image editors, evaluating the approach on two open-weight backbones.

Critical strengths

A useful benchmark and automated evaluation framework for the community. The dataset spans natural scenes, graphic designs, digital interfaces, and documents, bringing long text, length mismatch, and multi-region replacement into a unified evaluation. Combining OCR with visual assessment provides a useful basis for evaluating visual text editing more comprehensively than text accuracy alone.

Critical weaknesses
  1. GlyphPlan's necessity and independent benefits are not sufficiently established. The layout planner is central to the method, yet direct comparisons with leading proprietary vision-language models are missing, as are adequate comparisons with simpler approaches using strong models and structured prompts. The current experiments therefore do not establish whether dedicated planner fine-tuning and the additional glyph-processing pipeline are warranted.
  2. The English-only scope limits applicability. The data and experiments are restricted to English, leaving other scripts, line-breaking conventions, and mixed-language layouts untested. This limits confidence that the benchmark and method represent broader real-world text-editing needs.
Questions for authors
  1. Could the authors compare the fine-tuned planner with leading proprietary vision-language models or strong models without task-specific fine-tuning, while keeping the glyph renderer and image editor fixed? Is there evidence on planning quality, final editing performance, and cost that justifies dedicated fine-tuning?
  2. Why is the benchmark restricted to English, and what language-dependent assumptions are made in data construction or evaluation? Are there results on other languages or mixed-language inputs, or an explanation of the main limitations to extending the benchmark?
  3. Under the same requirements for preserving the source image's style and unedited content, can recent strong models with detailed structured prompts directly perform these text-replacement and reflow tasks? In which scenarios does GlyphPlan's additional pipeline provide clear benefits over this simpler alternative?
  4. Why was Qwen3-VL-235B-A22B-Instruct selected for text and layout annotation rather than newer leading vision-language models? Are there comparisons of annotation accuracy, agreement with human annotations, or cost that support this choice and establish that the annotations are sufficiently reliable for subsequent training and evaluation?
4 comments
AuthorsOpenReview ↗

Response to Reviewer xfwi

Thank you for your constructive feedback. In response, we provide controlled planner comparisons, clarify the English-only scope, evaluate structured prompting, and support our choice of annotation model with a human audit.

Q1. Does task-specific planner training help?

We replace only the planner, keeping the Arial renderer, with-glyph SFT editor, editing prompt, and sampling settings fixed within each backbone. All planners are evaluated on the same 2,002 images. Final scores are summarized below; complete results are provided in Appendix H.1, Tables 19–20:

PlannerSenseNova Final ↑Qwen Final ↑
GlyphPlan (trained)66.2853.19
GPT-6.1 Sol65.4051.82
Gemini 3.8 Flash64.8749.33
gemma-4-31B-it64.6249.12
Qwen3.6-27B (base)63.6247.54
Qwen3.5-9B63.1844.72
GPT-6 Luna61.7447.76

Here, SenseNova and Qwen denote SenseNova-U1.5-8B-MoT and Qwen-Image-Edit-2511. Compared with the Qwen3.6-27B VLM without task-specific fine-tuning, GlyphPlan improves Final by 2.66 points on SenseNova and 5.65 points on Qwen, highlighting the importance of training the planner for text replacement and layout adaptation. The other tested API-based and open-weight planners also achieve lower Final scores on both editors. GPT-6.1 Sol is the closest alternative and leads on a few individual metrics, but still trails GlyphPlan by 0.88/1.37 Final points. API planners incur recurring service charges, whereas our open-weight planner can be deployed locally.

Q2. Why English only?

We agree that multilingual, mixed-language, and translation-driven replacement are valuable extensions. Our current benchmark focuses on long-text replacement, multiple jointly edited regions, and adaptive reflow. Its data construction, experiments, and evaluation form a coherent framework that supports our conclusions within this scope. Multilingual editing adds a distinct research dimension, requiring coordinated work on data collection, language-specific annotation, model training, and validation of the evaluation pipeline. Completing these steps is beyond the rebuttal phase, but we appreciate this suggestion and regard multilingual extension as an important direction for future work.

Q3. Can a detailed prompt replace the extra pipeline?

Appendix H.2 tests a structured prompt combining detailed requirements for text accuracy and source-style/background preservation with step-by-step guidance for localization, layout planning, text removal, and target rendering (see Appendix H.3 for the complete prompt design). Full results from Table 21 are:

EditorSettingFinal ↑FG-NED ↓FG-Recall ↑Style ↑Layout ↑Background ↑BG-Text F1 ↑
SenseNovaStructured prompt64.5818.9658.5359.8383.4413.2950.72
SenseNovaSFT + GlyphPlan66.2818.5757.1073.0683.0516.3053.09
QwenStructured prompt26.9571.295.5849.4967.629.1617.85
QwenSFT + GlyphPlan53.1928.4638.2960.0275.3114.0630.32

Despite detailed requirements and step-by-step guidance, structured prompting does not match SFT + GlyphPlan on either editor. The Final gap is much larger on Qwen (26.24 points) than on SenseNova (1.70 points). The long prompt may overburden Qwen, whereas SenseNova's reasoning capability as a unified multimodal model may help it handle lengthy instructions.

Q4. Why Qwen3-VL-235B for annotation?

We selected Qwen3-VL-235B-A22B-Instruct. Its technical report [1] describes 30 million OCR pretraining samples and strong OCRBench/OCRBench-v2 results. We then directly audited the same 200 images with four annotation systems (Appendix C.3, Table 10):

Annotation systemHuman acceptanceHuman–human agreement
Qwen3-VL-235B (FP8-dynamic)72.0%94.0%
Qwen3.6-27B67.0%89.5%
Qwen3.8-27B56.5%91.0%
PaddleOCR (PP-OCRv5 server)47.0%96.5%

Two annotators independently assess whether each image has correctly aligned boxes without missing required boxes, then discuss and adjudicate disagreements to obtain the final acceptance labels. The reported acceptance rates use these adjudicated labels; human–human agreement is measured before adjudication and reaches 92.8% overall. Qwen235B ranks first in acceptance. Figure 11 shows similar performance on easy cases, but more accurate localization and coverage from Qwen235B on difficult cases. These results support its use for our annotation task. Moreover, our training data undergo VLM-based filtering and human review to remove inaccurate OCR annotations and reduce their impact on experimental results.

[1] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025.

Dear Reviewer xfwi,

Thank you for your time and constructive feedback. We have addressed your four questions in our rebuttal and summarize the main updates below:

  • Planner training (Q1). With the renderer and adapted editor fixed, GlyphPlan achieves the highest Final among seven planners on both backbones, improving over its unadapted base by 2.66/5.65 points (Appendix H.1, Tables 19–20).
  • English-only scope (Q2). We appreciate your suggestion and agree that multilingual editing is valuable. Our benchmark, experiments, and evaluation coherently address long-text, multi-region replacement and reflow. Multilingual extension requires additional data collection, annotation, and evaluation beyond this rebuttal phase; we regard it as important future work.
  • Structured prompting (Q3). Detailed, step-by-step prompts remain below SFT + GlyphPlan by 1.70 Final points on SenseNova and 26.24 on Qwen, supporting the benefit of our pipeline over this alternative (Appendix H.2, Table 21).
  • Annotation model (Q4). A human audit of four systems on the same 200 images finds Qwen3-VL-235B has the highest bounding-box acceptance rate (72.0%), supporting our choice of annotation model (Appendix C.3, Table 10).

As the discussion period ends in two days, we would greatly appreciate it if you could review these updates and let us know whether they address your concerns. If any point remains unclear, we would be glad to provide further clarification. Thank you again for helping us improve the paper.

Best regards,
The Authors

Dear Reviewer xfwi,

Thank you again for your time and valuable feedback. As only a few hours remain in the discussion period, we would sincerely appreciate it if you could kindly review our rebuttal and let us know whether your concerns have been adequately addressed. If so, we would be very grateful if you could consider updating your score accordingly. Should you have any remaining questions or concerns, we would greatly appreciate hearing from you at your earliest convenience and would be happy to address them promptly.

Thank you once again for your time and consideration.

Best regards,
The Authors

Reviewer_xfwiOpenReview ↗

Response

Thank you for your response. The rebuttal address most of my concerns and I will raise my score to 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Updates and Contributions

We thank all reviewers for their constructive feedback and thoughtful discussion. Their comments helped us strengthen the experimental evidence and clarify the scope of our claims. Shared questions concerned performance across difficulty levels and the value of task-specific layout planning; individual concerns further motivated analyses of evaluation reliability, annotation quality, alternative prompting, and font-style preservation. Below, we summarize our responses and the paper's contributions.

Updates addressing shared concerns

  • Performance across difficulty levels (Us3U, MGEx). We add analyses of all 14 main-table editors by cumulative target-text length, replacement-length mismatch, number of edited regions, and domain. All 14 editors score lower in the highest versus lowest text-load and region-count groups; contraction and expansion are harder than equal-length replacement on average. Representative failures illustrate text corruption, layout errors, background loss, and style changes (Appendix G, Tables 16–18 and Figures 12–16).
  • Value of planner training (xfwi, Us3U). We compare GlyphPlan with six alternative planners, including three proprietary APIs and three open-weight VLMs, while fixing the renderer and adapted editor. GlyphPlan achieves the highest Final on both backbones, improving over its unadapted base by 2.66 points on SenseNova and 5.65 on Qwen. These controlled comparisons support task-specific planner training (Appendix H.1, Tables 19–20).

Additional experiments and clarifications

  • Font-style preservation and interpretation (Us3U). We add handwritten/decorative examples and report Style improvements on both Handwritten Casual and Movie Poster for both editors in our follow-up. Attention analysis is complemented by DINOv2 and diffusion feature comparisons: source–output text-region similarity improves by 1.87–3.06%, with all four comparisons remaining significant after Holm correction. Gains are positive across all five reported transformer layers (Appendix I, Figures 17–18 and Tables 22–24). These results provide further evidence of source-appearance preservation and inform our explanation of the Style gains.
  • Structured prompting (xfwi). We test detailed prompts with step-by-step editing guidance. SFT + GlyphPlan achieves higher Final on both backbones, exceeding structured prompting by 1.70 points on SenseNova and 26.24 on Qwen (Appendix H.2, Table 21).
  • Annotation quality (xfwi). To assess our choice against newer alternatives, two annotators audit four annotation models, including newer VLMs and a dedicated OCR baseline, on the same 200 images. Qwen3-VL-235B achieves the highest bounding-box acceptance rate, 72.0%, supporting its selection (Appendix C.3, Table 10).
  • Copy-input control and human agreement (MGEx). We now evaluate unchanged source images using the full-image OCR and ReflowJudge pipeline. Copy-input obtains only 5.11% FG-Recall and 11.90/60 foreground credit, below 13 of 14 editors. Its Final is 51.55, ranking 11th among 15 methods, as unchanged images earn near-perfect Style, Layout, and background-related scores. Thus, source–target text overlap does not yield a high foreground score (Appendix B.6, Table 8). We also report 85.4% human–human agreement, providing context for ReflowJudge's 79.1% agreement with adjudicated labels (Appendix F.1, Table 12).

We also clarify the English-only scope and the additional data, annotation, and evaluation work needed for multilingual extension; the human study's 700 outputs, 2,100 output–dimension units, and unified image-level agreement rule; Background strictness and ongoing calibration; and GlyphPlan's adaptation requirements, downstream layout evaluation, and reliance on the editor for font appearance and curved text (xfwi, MGEx).

Key contributions

  • ReflowText: A full-image benchmark with 2,002 images and 9,197 replacement regions across four domains and 15 subcategories. It isolates long text, length mismatch, and multiple jointly edited regions as coupled challenges requiring adaptive reflow.
  • ReflowJudge: A human-validated visual evaluator that complements OCR metrics to assess text accuracy, font-style consistency, layout and reflow, and background preservation separately, enabling diagnosis beyond text recognition alone.
  • GlyphPlan: A modular approach that separates target-text layout planning from image editing through spatially aligned glyph conditions. Controlled experiments on two editor backbones show gains over matched no-glyph SFT and demonstrate the importance of learning to use the planned conditions.

We sincerely thank the reviewers for their constructive feedback and guidance. We have carefully conducted extensive additional experiments and provided point-by-point responses to every question and identified weakness. We hope that these efforts and clarifications address the reviewers' concerns and strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.