acceptodds
Under review as a conference paper at ICLR 2027

Document Post-Train for Image Generation

Abstract

Document page generation with text-to-image models requires accurate rendering of long passages and precise layout control, enabling automated creation of text-rich materials. However, existing models often suffer from local text defects and inaccurate layouts. To address these limitations, we introduce a post-training method covering both data and optimization. On the data side, we construct an automated pipeline turning real PDF pages into OCR-grounded dense captions with layout and visual features. Regarding optimization, we design legibility rewards based on VLM and region-level credit assignment specifically to address local text defects. Evaluations demonstrate that an open-source model trained with our method achieves commercial-level performance on our benchmark, ranking second only to GPT Image 2 overall, and this improvement also generalizes to other text-rich tasks. Our further analysis reveals the ineffectiveness of image-level optimization and OCR rewards, and shows that reinforcement learning reduces sampling needs while retaining diversity for test-time scaling. We believe these findings offer a practical path toward high-quality document generation.

Primary area:
generative models
Keywords:
document page generation, text-rich image generation, region-level reinforcement learning

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_Wu5LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses bilingual document-page generation, where long passages must remain legible while satisfying layout and appearance requirements. It introduces DocPage-1M, comprising 969,808 pages from 361,517 PDFs and averaging 1,372 visible characters per page. dots.mocr supplies regional text and boxes, while VLMs describe pictures and global design; refilling these references produces dense captions with content, coordinates, and typographic attributes. SenseNova-U1.5-8B-MoT-Preview is fully fine-tuned using flow matching on 70% document and 30% general data, then optimized with Flow-GRPO and MixGRPO-style sampling. A Gemini VLM assigns three-level regional legibility rewards, and group-relative regional advantages are mapped to spatially corresponding image tokens for local policy updates. On the authors’ 400-page bilingual DocPageBench, SFT+RL with their prompt enhancement reports Text/Layout/Aesthetics scores of 0.8260/0.8686/0.8425 (Avg. 0.8457), versus 0.8585 Avg. for GPT Image 2. BizGenEval Avg.(4) increases from 0.6785 for Base to 0.7068. Further experiments examine SFT initialization, page-level updates, OCR rewards, and best-of-N test-time selection, including RL@1 scoring 0.8480 versus 0.8507 for SFT@64.

Critical strengths
  1. The regional post-training recipe targets an important failure mode and is supported by multi-faceted analysis.
  2. The work combines a substantial supervision resource with practical analysis of inference cost.
Critical weaknesses
  1. The headline comparison with existing generators is confounded by prompt adaptation. In Table 2, SFT+RL uses Our PE, for which Gemini 3.1 Pro rewrites the request, completes missing content, and inserts coordinates when absent (Appendix C.3), whereas every external baseline uses Original requests. This choice is material: under Original inputs, SFT+RL obtains Avg. 0.6483 and Layout 0.7325, rather than 0.8457 and 0.8686 with Our PE. Its Original Layout is below GPT Image 2 (0.7542), Seedream-5.0-Pro (0.7396), and Qwen-Image-3-Pro (0.7384). Base also improves from 0.5845 with Original inputs to 0.6609 with Our PE, independently demonstrating the adapter’s effect. Consequently, the claims of ranking second only to GPT Image 2 and achieving the highest Layout do not establish generator superiority under a common input pipeline. This directly affects the paper’s principal empirical positioning.
  2. The experiments do not isolate the contribution of spatial credit assignment, and variability is unmeasured. Figure 6 compares the full method with a page-level variant, but Appendix D.3 explicitly notes that the variants also differ in loss reduction, including a window-length factor, and that these effects were not separately ablated. Moreover, Appendices D.1 and E.1 report only one training run per configuration. Thus, the observed 0.5559-versus-0.4570 result supports the complete regional update recipe over one page-level implementation, but does not establish how much benefit comes specifically from mapping advantages to local image tokens—the central algorithmic distinction from prior scalar-reward approaches. Without a mask-only control and an estimate of run variability, the paper’s mechanistic interpretation remains under-supported.
Questions for authors
  1. Can you evaluate all external baselines with the same saved Our PE captions, or otherwise provide a common prompt-adaptation budget, and report the resulting rankings? If such adaptation is unsuitable for some models, what is the justification for treating Original-versus-Our PE as the main comparison rather than the available Original-versus-Original results?
  2. Can you provide a control that removes only the spatial masks while holding regional advantages, loss normalization/reduction, window-length factors, and optimization budget fixed? Do the Figure 6 and Figure 7 conclusions persist across multiple training seeds, and what is the observed variability?
  3. What human or independent-judge validation supports the Gemini L1/L2/L3 legibility labels and DocPageBench’s VLM-based Layout and Aesthetics scores? In particular, does independent evaluation preserve the reported ordering of SFT+RL, SFT, OCR-reward RL, and page-level RL? Appendix E.3 currently acknowledges that the shared-judge analysis does not establish human preferences.
  4. What are the source collections and licenses for the 361,517 PDF files, what screening was performed for personal or sensitive content, and do the authors have redistribution rights for the page images and derived captions used in DocPage-1M and DocPageBench?
3 comments
AuthorsOpenReview ↗

Response Part 1

We thank the reviewer for the careful assessment and for recognizing the value of the regional post-training approach and its supporting analyses.

Q1: How do all external baselines rank when evaluated with the same PE descriptions?

A: We have now evaluated all eight external baselines under the same PE inputs. Their outputs were regenerated using exactly the same 400 saved PE descriptions as our model, giving all models the same text and layout information.

Model (all using the same PE)TextLayoutAestheticsOverall
GPT Image 20.84740.75990.98980.8657
Ours (SFT+RL)0.82600.86860.84250.8457
Seedream 5 Pro0.81770.75550.92650.8332
Nano Banana 20.81570.73820.91840.8241
Qwen Image 3 Pro0.79240.73140.90150.8084
HunyuanImage-30.35540.49390.53090.4600
FLUX.2-dev0.30170.44830.59240.4475
Qwen-Image-25120.37090.46870.44310.4276
Z-Image0.40520.48430.31210.4006

Each metric is averaged over the valid samples for each model.

Our findings:

  • Overall competitiveness is preserved under matched inputs. Our model ranks second among all nine models, behind only GPT Image 2, with an Overall score of 0.8457.
  • The layout advantage also persists. Our model achieves the highest Layout score, 0.8686, even when all baselines receive the same text and layout descriptions.

We will add this comparison and explicitly state that the ranking holds under the shared PE condition.


Q2: Does a matched control isolate spatial feedback, and do the ablation conclusions hold across training seeds?

A: To isolate the contribution of spatial feedback, the new control makes only one change to the optimization: regional advantages are pooled into a single value per page and applied to the same text regions. The SFT initialization, regional reward computation, text-region weighting, loss normalization and reduction, window-length factor, and 200-update training budget are all kept identical. Both variants are evaluated on the same 256 validation pages.

Feedback allocation (both at 200 updates)VLM legibility ↑L3 regions (%) ↑L1 regions (%) ↓
Uniform feedback within each page (new matched control)0.480839.5928.52
Spatially corresponding feedback (ours)0.555949.4124.27

Both settings cover the same 1,273 target regions, with three complete ratings per region. L1 and L3 denote the lowest and highest legibility tiers, respectively.

To assess training variability, we also repeated three configurations—regional VLM feedback, uniform VLM feedback within each page, and regional OCR feedback—with three training seeds (42, 123, and 2026) per configuration. For this comparison, we evaluate all RL runs at 100 updates on the same 256-page validation set.

Training feedback (all RL runs at 100 updates)VLM legibility ↑OCR ↑L3 regions (%) ↑L1 regions (%) ↓
SFT (shared step-0 reference)0.42390.861633.5431.26
Uniform VLM feedback within each page (matched control)0.4570 ± 0.03580.8846 ± 0.005237.16 ± 3.5631.16 ± 2.52
Regional OCR feedback0.4932 ± 0.00520.8855 ± 0.002339.70 ± 3.2228.17 ± 2.11
Regional VLM feedback (ours)0.5398 ± 0.01290.9053 ± 0.004844.65 ± 0.5924.85 ± 2.31

RL results are mean ± sample standard deviation across three training seeds; SFT is a single shared reference. Each run covers the same 1,273 target regions, with no missing VLM ratings.

Our findings:

  • Spatial feedback is beneficial under matched conditions and across the tested seeds. At 200 updates, spatially corresponding feedback improves mean legibility by 0.0751, increases L3 by 9.82 percentage points, and reduces L1 by 4.24 points. In the 100-update comparison, it also achieves higher VLM, OCR, and L3 values and a lower L1 proportion than uniform page feedback for every seed.
  • The reward-choice result is also consistent across the three seeds. Regional VLM feedback outperforms regional OCR feedback on all four metrics for every seed, with mean VLM/OCR gains of 0.0466/0.0198. This supports the benefit of VLM-based regional rewards for text quality across the tested training runs.

We will add the controlled comparison, the means and standard deviations, and the per-seed results to the revised manuscript.

AuthorsOpenReview ↗

Response Part 2

Q3: Do human or independent judgments validate the reported quality gains and regional legibility grades?

A: To validate aesthetics and legibility beyond the training judge, we added blinded evaluations of the aesthetic improvement from SFT to RL, text-rendering quality in two key ablations, and agreement with the L1/L2/L3 regional legibility grades. Each task contains 50 samples, each independently assessed by three annotators. The two ablation comparisons use models trained for 200 updates.

EvaluationSamplesResult
SFT+RL vs SFT: page aesthetics50 pairsPreference for SFT+RL: 73% [61%, 84%]
Spatial feedback vs uniform page feedback: text quality50 pairsPreference for spatial feedback: 73% [61%, 84%]
Regional VLM reward vs regional OCR reward: text quality50 pairsPreference for VLM reward: 86% [76%, 94%]
Regional grades vs annotator consensus50 regionsExact agreement: 72%; weighted κ = 0.669

Preference scores use consensus judgments, counting ties as one half; brackets indicate 95% confidence intervals. Regional consensus is the median of the three grades.

Our findings:

  • Independent judgments preserve the direction of the key comparisons. They favor RL over SFT in page aesthetics, spatially corresponding feedback over uniform page feedback, and VLM rewards over OCR rewards in text-rendering quality, supporting the main quality gains and the two ablation conclusions.
  • Independent annotations support the grading used in our legibility reward. Agreement with the original regional grades reaches 72%, with a weighted κ of 0.669. Only 1 of 50 regions (2%) exhibits a two-level disagreement between L1 and L3, providing additional support for the regional quality assessments underlying our reward.

We will include these evaluations and the annotation protocol in the revised manuscript.


Q4: What are the data sources and licenses, and how are screening and redistribution handled?

A: Our PDFs were collected from publicly accessible datasets and web sources for research and have been filtered through a sophisticated automated pipeline. We plan to release the model weights. Before publicly releasing the page images and derived captions from DocPage-1M and DocPageBench, we will conduct another round of rigorous screening and manual review and verify redistribution permissions to ensure that the released materials contain no personal or sensitive information and comply with the applicable licensing terms. We will document the main source collections, applicable licenses, and specific release scope in the revised manuscript's data statement.

AuthorsOpenReview ↗

Following Up on Our Response

Dear Reviewer Wu5L,

Thank you again for your careful assessment and helpful suggestions. We hope our response addresses the points you raised and would be grateful to hear whether any substantial issues remain.

Best regards,

The Authors

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

This paper studies document-page generation using text-to-image models to render long passages. It builds DocPage-1M, a large PDF dataset with OCR-grounded captions. Besides, it introduces a post-training recipe which follows a flow-matching SFT, followed by a RL with region-level VLM legibility reward and spatial credit assignment to address local text defects. The method is applicable to current open models and achieves good performance on DocPageBench.

Critical strengths
  • The dataset built by the authors has great value to this community. I think the dataset design is thoughtful, clearly effective to leverage VLM write descriptions over indexed references and then refilling OCR text.

  • The region-level credit assignment is the part I like most. Mapping the region's advantage onto the image token is effective to handle dense pages, compared with collapsing everything into one scalar. The ablation study also verify the large gain from this design.

  • The OCR-vs-legitibility finding is valuable to the community. It shows that the OCR rewards are the wrong target with supporting data points, and people should learn from this observation.

Critical weaknesses
  • Region geometry — overlaps and box-edge incompleteness — is unhandled in the data and unsafeguarded in the reward. The overlapped boxes/regions are never considered in the DocPage-1M dataset. Also, for the incomplete/cut words near the box edge, the author name this problem but only fix the other branch. For example, boxes are expanded 12% per side (App. B.2) to accommodate glyphs exceeding the box. App. E.1 then admits the cost — the margin admits "neighboring lines" — and the OCR reward is explicitly built to tolerate it. However, the VLM legibility reward has no such allowance to handle this issue.

  • The comparison is not fail. In table 2, the 8 baselines are evaluated with original request, compared with the proposed model with our PE, which uses a LLM to rewrite the original request into dense captions. I understand that the PE is part of the contribution, but I would like to see the baseline results under the PE. Until then, the abstract should state the ranking under matched inputs.

  • The LLM judges are almost the same family (gemini-3.8-flash) for reward, selector, and evaluation. This exist potential risk that the method development converge to a local optimal due to the judges' bias. Better involve multiple LLM judges or human evaluations to ensure the fairness.

Questions for authors
  • Compared with GPT Image 2, why do people use this method to fine-tuning open model? What's the reasons using the proposed method? There is not other comparison regarding to the latency, training efficiency, etc.

  • What do GPT Image 2, Qwen-Image-3-Pro, etc, score with our PE captions? This comparison is needed.

  • How sensitive is the three-level reward mapping (0.1, 0.5, 1.0)?

4 comments
AuthorsOpenReview ↗

Response Part 1

We thank the reviewer for recognizing the value of DocPage-1M, regional credit assignment, and the comparison between OCR and legibility rewards.

Q1: How are region overlaps and crop boundaries handled in the data, optimization, and reward evaluation?

A: We address this concern at three levels.

Overlap in the dataset. During data curation, we applied a comprehensive set of filtering rules to exclude low-quality pages with severe overlap. Our full-dataset analysis shows that only 0.58% of text boxes in DocPage-1M overlap another text box. We also believe it is important for the model to learn to handle reasonable, limited overlap in document layouts.

Feedback at overlapping positions. We thank the reviewer for raising this point and agree that this mechanism should be explained more clearly. In Eq. (4), the advantages of regions covering the same image token are combined using an inverse-area-weighted average. Each overlapping position therefore receives a single aggregated advantage. We will clarify this treatment in the revised manuscript.

Reward evaluation near crop boundaries. To assess the effect of crop boundaries on the VLM reward, we added a crop-margin sensitivity study on 100 matched requests (50 English and 50 Chinese). We evaluated the same 200 saved images from SFT and the 200-update RL checkpoint using crop margins of 0%, 6%, and 12%.

Crop marginSFTRLRL − SFT
0%0.38330.4794+0.0961
6%0.39040.4842+0.0938
12%0.38810.4699+0.0818

Scores are mean text-legibility rewards. All settings use the same images and target regions.

Our findings:

  • Mean scores remain close across the tested crop settings. The maximum variation across settings is 0.0071 for SFT and 0.0144 for RL, indicating that average legibility assessments are relatively stable under these crop-margin changes.
  • RL's text-quality advantage persists across crop settings. The gain over SFT ranges from 0.0818 to 0.0961, with positive paired 95% confidence intervals in all three settings, supporting the consistency of RL's legibility improvements across the tested crop extents.

We will add the dataset statistics and crop-margin analysis to the revised manuscript.


Q2: How do all external baselines rank when evaluated with the same PE descriptions?

A: We have now evaluated all eight external baselines under the same PE inputs. Their outputs were regenerated using exactly the same 400 saved PE descriptions as our model, giving all models the same text and layout information.

Model (all using the same PE)TextLayoutAestheticsOverall
GPT Image 20.84740.75990.98980.8657
Ours (SFT+RL)0.82600.86860.84250.8457
Seedream 5 Pro0.81770.75550.92650.8332
Nano Banana 20.81570.73820.91840.8241
Qwen Image 3 Pro0.79240.73140.90150.8084
HunyuanImage-30.35540.49390.53090.4600
FLUX.2-dev0.30170.44830.59240.4475
Qwen-Image-25120.37090.46870.44310.4276
Z-Image0.40520.48430.31210.4006

Each metric is averaged over the valid samples for each model.

Our findings:

  • Overall competitiveness is preserved under matched inputs. Our model ranks second among all nine models, behind only GPT Image 2, with an Overall score of 0.8457.
  • The layout advantage also persists. Our model achieves the highest Layout score, 0.8686, even when all baselines receive the same text and layout descriptions.

We will add this comparison and make the shared PE condition explicit when reporting the ranking in the abstract and main results.

AuthorsOpenReview ↗

Response Part 2

Q3: Do independent evaluations support the gains, given that the original judges belong to the same model family?

A: To validate aesthetics and legibility beyond the training judge, we added blinded evaluations of the aesthetic improvement from SFT to RL, text-rendering quality in two key ablations, and agreement with the L1/L2/L3 regional legibility grades. Each task contains 50 samples, each independently assessed by three annotators. The two ablation comparisons use models trained for 200 updates.

EvaluationSamplesResult
SFT+RL vs SFT: page aesthetics50 pairsPreference for SFT+RL: 73% [61%, 84%]
Spatial feedback vs uniform page feedback: text quality50 pairsPreference for spatial feedback: 73% [61%, 84%]
Regional VLM reward vs regional OCR reward: text quality50 pairsPreference for VLM reward: 86% [76%, 94%]
Regional grades vs annotator consensus50 regionsExact agreement: 72%; weighted κ = 0.669

Preference scores use consensus judgments, counting ties as one half; brackets indicate 95% confidence intervals. Regional consensus is the median of the three grades.

Our findings:

  • Independent judgments preserve the direction of the key comparisons. They favor RL over SFT in page aesthetics, spatially corresponding feedback over uniform page feedback, and VLM rewards over OCR rewards in text-rendering quality, supporting the main quality gains and the two ablation conclusions.
  • Independent annotations support the grading used in our legibility reward. Agreement with the original regional grades reaches 72%, with a weighted κ of 0.669. Only 1 of 50 regions (2%) exhibits a two-level disagreement between L1 and L3, providing additional support for the regional quality assessments underlying our reward.

We will include these evaluations and the annotation protocol in the revised manuscript.


Q4: Why fine-tune an open model instead of using GPT Image 2, and what is the generation latency?

A: Our model combines competitive document quality with low generation latency and the flexibility of local deployment and further adaptation. We summarize the average per-page generation time on DocPageBench below.

ModelAverage generation time
SFT (local)16.3 s
Ours (local)16.5 s
GPT Image 2 (API)83.8 s

Our findings:

  • Substantially lower generation latency than GPT Image 2. Our model averages 16.5 seconds per page on DocPageBench, compared with 83.8 seconds for the GPT Image 2 API, corresponding to approximately 80% lower latency. Together with its second-place Overall and best Layout scores under matched PE inputs (Q2), this offers a practical balance between document quality and generation efficiency.
  • Direct control over the model and deployment. Commercial services such as GPT Image 2 do not give users direct control over the underlying model architecture, weights, or any service-side prompt-enhancement pipeline. Users cannot independently modify or retrain the model on their own infrastructure, or deploy it locally. We will release our model weights to enable local deployment and further task-specific fine-tuning, allowing users to adapt the model to their own document styles, domains, and layout requirements.

We will add these timings and clarify the deployment benefits in the revised manuscript.

AuthorsOpenReview ↗

Response Part 3

Q5: How sensitive is the three-level reward mapping (0.1, 0.5, 1.0)?

A: We agree that the choice of reward values warrants a sensitivity check. We varied the middle-level reward to 0.3 and 0.7 while keeping the other two values fixed. All three variants use the same SFT initialization, training seed, and optimization settings, and are compared after 60 updates on the same 256 validation pages.

Training reward mappingValidation reward ↑L3 regions ↑L1 regions ↓
SFT before RL0.416334.41%34.80%
(0.1, 0.3, 1.0)0.454838.57%31.66%
(0.1, 0.5, 1.0), default0.533844.78%24.27%
(0.1, 0.7, 1.0)0.499340.53%27.10%

For a consistent comparison, validation rewards are computed using the same reference mapping (0.1, 0.5, 1.0). L3/L1 proportions do not depend on the numerical mapping.

Our findings:

  • Quality gains persist across all tested mappings. All three variants improve validation reward, increase the proportion of L3 regions, and reduce L1 regions relative to SFT, showing that the gains are not confined to one exact reward assignment.
  • The default mapping performs best among the tested settings. It achieves the highest validation reward and L3 proportion, together with the lowest L1 proportion, supporting the reward mapping used in the paper. This may reflect a balanced incentive for both making unreadable text readable and refining readable text into clean rendering. A lower middle reward gives less credit for reaching L2, while a higher one reduces the additional reward for reaching L3.

We will include this sensitivity analysis in the revised manuscript.

AuthorsOpenReview ↗

Following Up on Our Response

Dear Reviewer XrBF,

Thank you again for your positive assessment and constructive suggestions. When you have a chance, please let us know whether the added analyses address your questions or if there is anything else we can clarify.

Best regards,

The Authors

Reviewer_gg5SOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates document generation using text-to-image generative models. Specifically, it addresses the challenges of generating long passages and achieving precise layout control. Its contributions consist of two parts: (i) DocPage-1M, a dataset containing 970,000 real bilingual PDF pages with dense captions, and (ii) VLM-based legibility rewards and region-level credit assignment to reduce local text defects. In experiments, the proposed method achieves strong results and demonstrates generalization on BizGenEval. The paper also analyzes the roles of SFT initialization, regional optimization, VLM rewards, and RL strategies. These findings validate the overall motivation and methods.

Critical strengths
  • Valuable problem and dataset.

The problem is valuable and well-motivated, as it exists in current cutting-edge generation models. The proposed DocPage-1M dataset approaches million-scale, has better resolution (Table 1), and is built from real PDFs.

  • Clean decomposition of the problem.

The paper analyzes data, reward, and credit assignment through separate ablations, and the narrative and experiments map one-to-one.

  • Comprehensive and honest appendix.

The paper provides comprehensive evaluation reports and intermediate results in the appendix, covering single-run training, missing evaluations, evaluation coverage, optimization, and judge-related issues. This reduces the verification burden.

  • Inspiring analysis of RL vs. test-time scaling.

The paper analyzes whether RL can substitute for or complement best-of-N selection, and quantifies this (RL@1 ≈ SFT@64).

Critical weaknesses

Confounding factors in the ablation study. The main ablation changes several factors at once. Appendix D.3 states that regional and page-level training differ in text-region weighting, loss normalization, and a window-length factor. Failed judge calls can also affect the two methods differently. Therefore, Figure 6 supports the full regional training setup but does not establish whether matching rewards to the correct spatial regions causes the improvement. This is my main concern.

Prompt adaptation makes the cross-model comparison difficult to interpret. If I understand correctly, the proposed PE completes the text and adds layout information, increasing the final model’s score from 0.6483 to 0.8457. Other open-source models mainly receive the original requests, indicating differences in input processing. Providing baselines with equivalent text and layout information would strengthen the comparison.

No human evaluation. Human evaluation would help validate the aesthetic gains from SFT to RL (0.7525 → 0.8425).

Validation of text quality. Related to the point above, the legibility ablations use the same VLM that supplies the training reward. The main benchmark uses a different judge, but agreement between human judgments and the legibility scores has not been established.

Minor concerns:

  1. The gap between the English and Chinese aesthetics scores (0.7818 vs. 0.9033) is large and unexplained.
  2. The choice of update 560 needs clarification. Table 15 shows that updates 600 and 700 score higher, and the sweep is conducted on the test benchmark.
Questions for authors

I use GPT models to read and analyze the key contributions of this paper in comparison with previous papers. I ask critical questions and request exact references or clarifications from the paper. I draft all final reviews, recommendations, weaknesses, and strengths myself, using GPT only to correct grammatical errors in those drafts.

3 comments
AuthorsOpenReview ↗

Response Part 1

We thank the reviewer for recognizing the value of the dataset, the experimental analysis, and the findings on RL and test-time scaling.

Q1: Does the ablation isolate the effect of assigning feedback to the correct spatial regions?

A: To isolate the contribution of spatial feedback, the new control makes only one change to the optimization: regional advantages are pooled into a single value per page and applied to the same text regions. The SFT initialization, regional reward computation, text-region weighting, loss normalization and reduction, window-length factor, and 200-update training budget are all kept identical. Both variants are evaluated on the same 256 validation pages.

Feedback allocation (both at 200 updates)VLM legibility ↑L3 regions (%) ↑L1 regions (%) ↓
Uniform feedback within each page (new matched control)0.480839.5928.52
Spatially corresponding feedback (ours)0.555949.4124.27

Both settings cover the same 1,273 target regions, with three complete ratings per region. L1 and L3 denote the lowest and highest legibility tiers, respectively.

Our findings:

  • Spatial feedback remains beneficial under matched optimization conditions. Assigning feedback to corresponding regions improves mean legibility by 0.0751, increases the proportion of L3 regions by 9.82 percentage points, and reduces the proportion of L1 regions by 4.24 points. This supports the value of preserving the correspondence between local quality differences and spatial locations when optimizing document text quality.

We will add this controlled comparison and its configuration to the revised manuscript.


Q2: How do the baselines perform when given the same PE descriptions as our model?

A: We have now evaluated all eight external baselines under the same PE inputs. Their outputs were regenerated using exactly the same 400 saved PE descriptions as our model, giving all models the same text and layout information.

Model (all using the same PE)TextLayoutAestheticsOverall
GPT Image 20.84740.75990.98980.8657
Ours (SFT+RL)0.82600.86860.84250.8457
Seedream 5 Pro0.81770.75550.92650.8332
Nano Banana 20.81570.73820.91840.8241
Qwen Image 3 Pro0.79240.73140.90150.8084
HunyuanImage-30.35540.49390.53090.4600
FLUX.2-dev0.30170.44830.59240.4475
Qwen-Image-25120.37090.46870.44310.4276
Z-Image0.40520.48430.31210.4006

Each metric is averaged over the valid samples for each model.

Our findings:

  • Overall competitiveness is preserved under matched inputs. Our model ranks second among all nine models, behind only GPT Image 2, with an Overall score of 0.8457.
  • The layout advantage also persists. Our model achieves the highest Layout score, 0.8686, even when all baselines receive the same text and layout descriptions.

We will add this comparison and state the matched-input ranking explicitly in the revised manuscript.

AuthorsOpenReview ↗

Response Part 2

Q3: Do human judgments support the aesthetic improvements and the regional legibility grades?

A: To validate aesthetics and legibility beyond the training judge, we added blinded evaluations of the aesthetic improvement from SFT to RL, text-rendering quality in two key ablations, and agreement with the L1/L2/L3 regional legibility grades. Each task contains 50 samples, each independently assessed by three annotators. The two ablation comparisons use models trained for 200 updates.

EvaluationSamplesResult
SFT+RL vs SFT: page aesthetics50 pairsPreference for SFT+RL: 73% [61%, 84%]
Spatial feedback vs uniform page feedback: text quality50 pairsPreference for spatial feedback: 73% [61%, 84%]
Regional VLM reward vs regional OCR reward: text quality50 pairsPreference for VLM reward: 86% [76%, 94%]
Regional grades vs annotator consensus50 regionsExact agreement: 72%; weighted κ = 0.669

Preference scores use consensus judgments, counting ties as one half; brackets indicate 95% confidence intervals. Regional consensus is the median of the three grades.

Our findings:

  • Independent evaluation supports the main results and key ablations. Blinded judgments favor RL over SFT in page aesthetics, spatially corresponding feedback over uniform page feedback, and VLM rewards over OCR rewards in text-rendering quality.
  • Independent annotations support the grading used in our legibility reward. Agreement with the original regional grades reaches 72%, with a weighted κ of 0.669. Only 1 of 50 regions (2%) exhibits a two-level disagreement between L1 and L3, providing additional support for the regional quality assessments underlying our reward.

We will include these evaluations and the annotation protocol in the revised manuscript.


Q4: Why are Aesthetics scores lower for English than for Chinese?

A: To better understand the English–Chinese aesthetics gap, we examined the reference-page text volume and language-specific Aesthetics scores on DocPageBench.

LanguageAverage characters per pageSFTSFT+RLGain
English1,995.40.68280.7818+0.0990
Chinese653.60.82230.9033+0.0810

Character counts are computed from reference-page OCR after removing formatting and whitespace.

Our findings:

  • Text-volume differences help contextualize the gap. English pages contain approximately 3.05 times as many characters on average. This greater text volume may partly explain the lower Aesthetics scores, as maintaining clean and consistent text rendering becomes more demanding.
  • RL benefits both languages and narrows the gap. RL improves Aesthetics in both languages, with a larger gain in English, reducing the language gap from 0.1395 to 0.1215.

We will add this analysis to the revised manuscript.


Q5: Why was update 560 selected when updates 600 and 700 score higher on DocPageBench?

A: We selected the checkpoint at RL update 560 based on Gemini reward on the validation set, before inspecting its DocPageBench scores. The higher DocPageBench scores at updates 600 and 700 in Table 15 were not used to select the reported main checkpoint. We will make this selection procedure explicit in the revised manuscript.

AuthorsOpenReview ↗

Following Up on Our Response

Dear Reviewer gg5S,

Thank you again for your thoughtful review. We would appreciate your thoughts on our response and the additional results, and would be happy to clarify any remaining concerns.

Best regards,

The Authors

Comments on the paper

2 comments

Dear Reviewers,

The authors have put considerable effort into addressing your comments, and the discussion phase is the best opportunity to resolve any remaining questions. Could you please review their rebuttal and let us know whether it changes your assessment? If some points are still unresolved, I encourage you to raise them with the authors so they have a chance to clarify. Your engagement is essential for reaching a fair and well-informed decision.

Thank you for your time and contribution.

Best regards,

AC

AuthorsOpenReview ↗

Summary of Contributions and Responses to Reviewer Concerns

Dear Area Chair and Reviewers,

We thank the reviewers and chairs for their constructive feedback. Our paper addresses a practical challenge: enabling open-source image models to generate full document pages with high-quality text and accurate layouts. We contribute training data built from nearly one million real English and Chinese pages, together with a reinforcement-learning method that evaluates the readability of individual text regions and directs training feedback to the corresponding image regions. This allows training to target local text defects within a page.

Our rebuttal provides new evidence addressing the reviewers’ main concerns:

  1. The reported ranking holds when every model receives the same information. Reviewers were concerned that our model benefited from more detailed prompts. We reran all eight external baselines with exactly the same text and layout descriptions as our model. It remains second overall, behind only GPT Image 2, and first in layout on our document-generation benchmark.

  2. Local feedback improves text quality under matched training conditions. Our new control combines the regional feedback into one value and applies it to the same text regions, keeping all other optimization settings fixed. After 200 updates, local feedback raises the proportion of regions receiving the highest legibility grade from 39.59% to 49.41%. A separate 100-update comparison shows the same advantage in all three training seeds.

  3. Human judgments support the quality gains and the key design choices. Blinded human evaluation favors reinforcement learning over supervised training in page aesthetics (73% preference), local over uniform feedback in text quality (73%), and legibility-based over OCR-based rewards in text quality (86%). Regional legibility grades also agree with human ratings in 72% of cases (weighted κ = 0.669). Each task uses 50 samples and three annotators; preference scores count ties as half.

We also addressed questions about overlapping regions and reward sensitivity: we added dataset-wide overlap statistics, clarified how feedback is combined at overlapping positions, and showed that the legibility gains persist across the tested crop margins and reward values.

We will incorporate these experiments and clarifications into the revised manuscript and welcome any remaining questions.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.