acceptodds
Under review as a conference paper at ICLR 2027

Panel2Page: Structured Visual Narrative Generation with Omni-Panel Alignment

Abstract

Structured visual generation requires models to satisfy hierarchical constraints beyond global text-image alignment, including global composition, local semantic correspondence, and cross-panel relational consistency. We study this challenge through text-to-comic page generation, where a model directly synthesizes a complete comic page from a structured prompt specifying layouts, ordered events, recurring characters, textual elements, and cross-panel relations. We propose Panel2Page, a two-stage framework that improves structured correspondence by separating global composition learning from hierarchical semantic alignment while preserving single-pass generation. Page-Level Supervised Fine-Tuning first learns a global structural prior from complete prompt-page pairs, capturing layout organization, reading order, and page-level composition. We then introduce Omni-Panel Alignment (OPA), a training-time strategy that regularizes unified page predictions using detached localized predictions constructed from complementary text-aware, panel-event, and cross-panel views, without introducing additional inference modules. To enable structured supervision, we construct a hierarchical dataset of approximately 20,000 manually verified prompt-page pairs and a benchmark of 320 structured prompts covering diverse layouts and narratives. Extensive experiments demonstrate that Panel2Page achieves the best overall performance among the evaluated full-page and panel-wise generation methods, showing the complementary benefits of global structural learning and hierarchical semantic alignment.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
structured visual generation, text-to-comic generation, hierarchical semantic alignment

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7NEcOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on generating complete comic pages from structured text prompts. The authors find that existing methods struggle to jointly preserve page layouts, panel events, textual fidelity, and consistency across panels. They therefore propose Panel2Page, which combines page-level supervised fine-tuning with Omni-Panel Alignment to align text, panel events, and cross-panel relationships. To validate their method, they conduct several experiments, including comparisons on a 320-prompt benchmark, alignment ablations, and a 25-participant user study. The experimental results demonstrate that Panel2Page achieves the highest overall score among the evaluated methods.

Critical strengths
  1. This paper proposes Panel2Page, which generates a complete comic page in a single pass from a structured prompt.

  2. To align text, panel events, and cross-panel relationships, the authors propose Omni-Panel Alignment, which adds no extra cost at inference.

  3. They construct a manually verified dataset of about 20,000 prompt-page pairs, providing hierarchical supervision for structured comic generation.

Critical weaknesses
  1. The paper emphasizes layout organization as a core requirement of comic pages and describes its benchmark as covering "diverse layouts." However, the layout is given in the prompt and restricted to four regular grids, and OPA constructs its views from these predefined panel boxes. It is therefore unclear whether the method extends to the irregular or variable-size panels common in real comics, which limits the generality of the claimed structured page generation.

  2. The paper treats cross-panel consistency as a core requirement and credits OPA's cross-panel views with improving it. However, consistency is assessed only within the VLM-judged Event score, which also covers reading order, temporal progression, and causal relations, and through broad user preferences. No dedicated measure isolates whether recurring characters actually remain consistent across panels.

  3. The 320-prompt benchmark is listed as a contribution, but its construction is not described. The paper does not state whether its prompts share the templates and LLM refinement used for the training prompts or whether they were checked for overlap. Since baselines see this structured format zero-shot, Panel2Page's advantage in Table 1 may partly reflect in-distribution prompts.

  4. The Abstract claims the best overall performance among the evaluated methods. Yet Panel2Page leads Gemini 3.1 Flash Image Preview by only 0.04 in Overall score (3.9830 vs. 3.9414) and trails it on Semantic and Page Quality (Table 1). Such a marginal lead does not clearly support the headline claim.

  5. The Event and Page Quality scores in Tables 1 and 2 lack explicit computation rules. The judge prompt in Appendix D.2 returns several component scores but does not specify how they are combined into these metrics, which limits reproducibility.

  6. Section 2 omits "DreamingComics: A Story Visualization Pipeline via Subject and Layout Customized Generation using Video Models" (CVPR 2026), a closely related comic pipeline that combines page-level layout planning with character and style consistency through multi-stage generation. The experiments also include no comparison with it, so the advantage of unified page generation over such a decomposed pipeline remains untested.

Questions for authors
  1. Could the authors evaluate Panel2Page on irregular layouts, such as variable-size or non-rectangular panels, and explain how OPA's views would be constructed in such cases?

  2. Could the authors report dedicated cross-panel consistency metrics, such as the character (CIDS) and style (CSD) self-similarity metrics from ViStoryBench ("ViStoryBench: Comprehensive Benchmark Suite for Story Visualization," CVPR 2026), computed across panels? Reporting them for the cross-panel ablation in Table 2 would also clarify the contribution of that view.

  3. Could the authors describe how the benchmark prompts were created and separated from the training data? Results on independently sourced structured prompts would help assess generalization.

  4. Could the authors demonstrate that the improvement over Gemini 3.1 Flash Image Preview is statistically significant?

  5. Could the authors provide the exact formulas and component weights for Event and Page Quality? Please specify how the scores are aggregated across samples.

  6. Could the authors discuss how Panel2Page differs from DreamingComics and compare it with DreamingComics on the same benchmark?

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

This paper presents Panel2Page, a two-stage framework for generating complete comic pages from structured prompts. It first fine-tunes qwen on full-page examples, then applies panel alignment, which aligns regions of full-page denoising predictions with detached predictions from the same model conditioned on localized text, panel events, and cross-panel relations. They construct around 20k prompt–page pairs and a 320-prompt benchmark. Comparisons, component ablations, and a 25-participant preference study show improvements, particularly in cross-panel narrative consistency.

Critical strengths

This paper proposes a training-time method for improving hierarchical semantic correspondences in full-page comic generation. OPA uses independent local denoising predictions from the same model as consistency targets for corresponding regions in the full-page prediction. These local branches are used only during training and do not introduce additional inference-time branches or sampling steps compared to the standard full-page generation process of the backbone network. The ablation experiments in Table 2 show that performance gradually improves from page-level SFT at ~3.79 to each individual viewpoint, ultimately reaching that of the full OPA ~3.98, with each viewpoint primarily improving the dimension it targets. These results demonstrate that the sources of performance improvement are relatively clear.

Critical weaknesses

1. the experimental comparison does not adequately cover the contemporary baselines. There are two types of baselines considered : one is panel-wise or storyboard methods, and the other is general-purpose text-to-image models (e.g., Qwen-Image, Gemini); neither is specifically designed for generating full-page comics with text. MangaFlow (Wang et al, cited but not compare) as well as DreamingComics (Patrick et al.), all directly address the generation of multi-panel or full-page comics but were not included in the comparison. Adding these pipeline could be more convincing to assess whether Panel2Page improves over the current state of the art.

2. The comparison between Page-Level SFT and SFT + OPA may benefit from a matching training budget. Since the OPA model undergoes an additional training phase on the SFT checkpoints, some of the improvement may stem from the continued training itself. It would be useful to compare against continued SFT trained for the same number of additional steps without OPA.

3. The writing needs substantial revision, with several inconsistencies and redundancies. For example, in the user study (L467–471), two consecutive sentences beginning with “In particular” make the same point about Panel2Page’s largest advantage.

Questions for authors
  1. Does the Page-Level SFT baseline in Table 2 receive the same total training epochs and learning-rate schedule as SFT + OPA?
  2. What are the “four narrative categories” in the benchmark? I could not find them defined in either the main paper or the appendix.
  3. Can you compare with MangaFlow, MangaDiffusion, and DreamingComics, or explain why direct comparison is not feasible?
Reviewer_1m6oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Panel2Page generates complete comic pages from structured prompts using page-level supervised fine-tuning followed by Omni-Panel Alignment (OPA). OPA aligns full-page predictions with detached local predictions from the same model across text, panel-event, and cross-panel views, without extra inference branches. The authors filter approximately 100,000 synthetic pages into 20,000 training pairs and evaluate on 320 prompts. Automatic evaluation, component ablations, and a 25-participant preference study support improved structured generation, particularly textual fidelity and event consistency.

Critical strengths
  1. The paper addresses a useful challenge, generating an entire comic page while keeping its text, characters, and story consistent. Combining whole-page and local supervision is a sensible approach to this problem.
  2. Comparisons, component ablations, and a user study provide several forms of evidence that the method helps, particularly with narrative consistency. The training changes also preserve the original inference process, which makes the approach practical to use.
Critical weaknesses
  1. OPA's contribution beyond continued page-level training is not fully isolated. Table 2 provides useful component ablations, but Appendix A.1 (lines 627–629) states that Page-Level SFT is trained for one epoch and OPA adds another epoch. The paper does not report a corresponding baseline that continues page-level training alone. Consequently, the comparison cannot fully distinguish the benefit of the proposed alignment from that of further training on comic pages. A control starting from the same SFT checkpoint, with the same data, update count, and learning-rate schedule but without OPA, would make the method's contribution clearer.
  2. The evaluation leaves the method's applicability to more varied comic layouts unclear. The benchmark contains 320 prompts covering four regular-grid layouts. Appendix E.4 (lines 960–964) also states that region construction assumes regular panel boxes and that irregular, overlapping, and non-rectangular panels have not been validated. This matters because OPA uses local and cross-panel regions to define its alignment targets; more flexible layouts may require changes to how these regions are constructed. The current results support the method within the tested layouts, but do not establish how broadly its alignment design applies. Clarifying these dependencies, with a small additional layout test where supported, would strengthen the evaluation.
Questions for authors
  1. Could the authors continue training the same SFT checkpoint using only the page-level denoising loss, matching OPA's training data, update count, and learning-rate schedule? How much of OPA's improvement remains against this control?
  2. Which parts of OPA's region construction depend on regular-grid layouts, and what changes would be needed for irregular or overlapping panels? If the current implementation supports other layouts, could a small evaluation show whether the alignment gains carry over? I would be happy to raise my score if the authors satisfactorily address these concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.