Pith. sign in

REVIEW 5 major objections 4 minor 52 references

The paper claims a six-agent pipeline, anchored by a style vector extracted from the core poster, can turn a raw dramatic script into a unified, interactive multi-page theater program set without retraining an image generator.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 11:11 UTC pith:VGUPM25A

load-bearing objection Novel domain and benchmark, but the circular VLM-as-a-Judge evaluation means the claimed superiority over baselines is not established. the 5 major comments →

arxiv 2607.19947 v1 pith:VGUPM25A submitted 2026-07-22 cs.CV

ETPDesigner: Multi-Agent Orchestration for Interactive Multimodal Electronic Theater Program

Mengtian Li , Xinru Guo , Xiaoru Lin , Xiao Rong , Zhifeng Xie , Chaofeng Chen This is my paper
classification cs.CV
keywords Electronic Theater ProgramMulti-Agent CollaborationStylistic ConsistencySelf-RefinementMultimodal InteractionPoster GenerationText-to-ImageBenchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that generating an Electronic Theater Program — a multi-page set of posters, information pages, and character portraits for a stage production — is a task that a system of specialized AI agents can perform end-to-end from a raw script. Its central move is to generate one 'core poster' first, extract a global style vector from it, and force every subsequent page (readable backgrounds, character portraits) to inherit that style. The authors argue this style-anchor mechanism is what prevents the style drift that plagues direct text-to-image generation, and they support it with a curated benchmark, automated judge scores, and a small human study. They further claim the same assets can be animated, given voices, and connected to persona-grounded dialogue, turning the static publication into an interactive companion. A sympathetic reader would care because it offers a template for automating any multi-page, style-constrained design task, not just theater programs.

Core claim

On the paper's own terms, the discovery is that 'theatrical abstraction' and cross-page consistency can be engineered without fine-tuning by decomposing the job into six cooperating agents: a semantic analyst that turns the script into themes and character profiles, an art director that formulates visual strategy, a generator, a critic that judges and refines outputs against retrieved professional posters, a style extractor, and a layout compositor. The load-bearing step is the global style anchor: after the critic-guided loop settles on the final core poster, the style extractor distills its artistic style and dominant color palette into an immutable prior that conditions all background and

What carries the argument

Global style anchor: a two-part visual prior (artistic style descriptor plus dominant color palette) distilled by the Style Extractor agent from the finalized core poster. It is the mechanism that enforces aesthetic uniformity: every subsequent background and character portrait is generated with that prior hard-conditioned into its prompt, so functional pages (which must have negative space for text) and character pages (which must be expressive) remain visually unified with the poster. The paper's ablations show this anchor, not the generation model itself, is responsible for the reported cross-asset consistency.

Load-bearing premise

The whole reported superiority rests on the automated critic's 1–10 ratings of theatrical aesthetics, communication efficacy, and thematic fidelity being faithful proxies for what human designers and audiences value—and on the iterative loop that pushes those ratings up actually improving the output rather than just gaming the judge.

What would settle it

Present a blind panel of professional theater designers with a set of competing final programs: the one selected by the automated judge, and a lower-scoring candidate from the same run. If experts do not prefer the judged-selected one at a statistically clear rate, the judge-based selection and the critic loop are not the source of the claimed quality.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • From one script, the full program package—poster, readable synopsis background, and character portraits—can be produced without retraining the underlying text-to-image model.
  • The style anchor ties all pages together: as long as the core poster is right, backgrounds and portraits inherit its look, which means improving the core poster is the highest-leverage improvement to the whole set.
  • Asset-level decoupling plus deterministic layout avoids garbled typography, so dense professional text such as cast lists and schedules can be rendered without bespoke text-generation.
  • The interactive layer (animated portraits, synthesized voices, persona-grounded dialogue) is a direct byproduct of the generated assets, so the same pipeline yields both a static program and a conversational character.
  • Removing any of the four identified components (expert mapping, critic loop, style extractor, asset composition) measurably degrades the reported quality, which the paper treats as evidence that the multi-agent decomposition is doing real work.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The orchestration pattern should transfer to any multi-page design artifact with one hero visual and several subordinate functional pages—festival brochures, exhibition catalogs, game art books—where a single style anchor would play the same unifying role.
  • Because the whole program inherits the core poster's style, a cheap practical upgrade is letting a human pick or lightly edit the core poster before the rest of the pipeline runs; the paper does not explore this human-in-the-loop variant.
  • The talking-character layer suggests an untested accessibility use case: an audio-interactive program could serve visually impaired theatergoers, since the same pipeline already produces voice and persona.
  • A stress test worth running: feed the system one script twice with different random seeds; if the two resulting programs differ wildly in style, the style anchor is under-constrained by the script and the 'faithfulness' claim needs qualification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes ETPDesigner, a training-free multi-agent framework that synthesizes electronic theater programs (ETPs) from raw dramatic scripts. The system chains a semantic analyst, an art director, a text-to-image generator, a VLM critic, a style extractor, and a layout compositor; the core poster is generated via an iterative RAG-enhanced Tree-of-Thoughts loop in which a VLM judge scores the output on theatrical aesthetics, communication efficacy, and thematic fidelity. A global style anchor is then extracted and used to generate style-consistent backgrounds and character portraits, which are programmatically assembled. The authors also introduce ETP-Pro, a benchmark of 100 curated programs, and report VLM-as-a-judge scores, objective consistency metrics (CLIP-I, DINOv2, OCR text accuracy), a small human study, and ablations to support the claim that ETPDesigner outperforms existing T2I and general multi-agent baselines.

Significance. If the evaluation is valid, the paper would make a useful contribution by demonstrating that a modular, training-free multi-agent pipeline can produce aesthetically coherent and semantically faithful multi-page theatrical publications, and by providing a domain-specific benchmark for this task. The global style-anchor mechanism and the decoupled asset-generation strategy are reasonable design choices. The construction of ETP-Pro is valuable as a new resource, and the interactive demonstration (animation, speech, persona-grounded dialogue) is an ambitious extension. However, the central numerical claims rest on an evaluation design that is circular in important ways; these concerns are load-bearing and must be addressed before the claimed superiority can be accepted.

major comments (5)
  1. [§3.2, Eq. (4); §5.1, Table 1] The evaluation is circular with respect to the Critic agent. During generation, A_crit scores every candidate poster on exactly the same three dimensions and eight sub-metrics (Theatrical Aesthetics, Communication Efficacy, Thematic Fidelity) that are later used as the primary VLM-as-a-judge evaluation metrics, and Eq. (4) selects the final poster as the argmax of this score. The proposed system is therefore explicitly optimized against the evaluation rubric, while all baselines and several ablations are not. The reported margins may reflect overfitting to this rubric rather than genuine quality differences. To establish the claim, the authors should (a) use an evaluation VLM that was never used during generation, (b) report the correlation between the VLM scores and human judgments, and (c) show that baselines are not disadvantaged by the absence of critic-style optimization.
  2. [§3.2, §4, §5.1] The RAG module queries ETP-Pro to retrieve reference posters during generation, and ETP-Pro is also the benchmark on which all results are reported. The paper describes no train/test split and no exclusion of the target script's own professional assets from the retrieval corpus. Because each ETP-Pro entry is 'strictly' paired with a source script, retrieval can return the ground-truth poster for the very script being generated, directly influencing strategy pruning and visual refinement. Even when the exact match is not retrieved, using the evaluation distribution as the retrieval corpus inflates scores. The authors must re-evaluate on a held-out set that is disjoint from the retrieval database, or at minimum demonstrate that retrieved references exclude the target script and report results after such exclusion.
  3. [Table 1 vs. Table 2] The ablation rows are internally inconsistent. In Table 1, the row 'Ours w/o Asset Comp' is identical to 'Ours' on every one of the ten VLM sub-metrics, yet Table 2 reports that removing asset composition drops DINOv2 from 0.684 to 0.528 and CLIP-I from 0.857 to 0.814. A VLM judge that cannot distinguish a DINOv2 drop of this magnitude is insensitive to exactly the structural property the paper claims to improve. Similarly, 'Ours w/o Style Extr' is identical to 'Ours' on all poster sub-metrics and only changes the whitespace score. This requires explanation: either the ablations were not executed as described, or the VLM metrics are too coarse to support the paper's claims. The authors should re-run the ablations and, if the reported numbers are correct, discuss why the VLM is insensitive to large objective changes.
  4. [§5.4, Fig. 5] The human evaluation is too thin to validate the automated metrics. It reports preference-level comparisons for 49 participants across four dimensions but provides no per-dimension numeric scores, no inter-rater agreement, no significance tests, and no correlation between the VLM-as-a-judge scores and human preferences. Since the paper's main quantitative claims—Theatrical Aesthetics, Communication Efficacy, Thematic Fidelity—are VLM scores, the authors must demonstrate that those scores are valid proxies for human judgment. Please provide correlation coefficients (e.g., Spearman) between the VLM judge and human ratings, with confidence intervals, and report statistical significance for the human-preference results.
  5. [§5.1, baselines] The baseline comparison is confounded by asymmetric optimization. The proposed system includes a VLM critic that iteratively optimizes outputs on the same dimensions later used as evaluation metrics, whereas the T2I baselines are only given an LLM recaption step and the AutoGen baseline is configured with generic agents. Thus Tables 1 and 2 partly compare a metric-optimized system against unoptimized pipelines. At minimum, the authors should run the baselines with the same number of refinement iterations and the same critic feedback, or justify why this is not possible, and report the performance of baselines under their own self-refinement loops.
minor comments (4)
  1. [Figure 4 caption] The caption lists 'PosterCraft' as a compared method, but §5.1 defines baselines as Nano Banana Pro, Seedream 4.5, and AutoGen. This inconsistency should be corrected or PosterCraft should be included in the experimental setup.
  2. [§4] The text says '100 unique electronic program' — should be 'programs'. Also, the paper does not describe any train/validation/test split for ETP-Pro, which is critical given the RAG-based method.
  3. [§5.1] The ETP-Pro statistics state 37 musicals, 63 spoken dramas, 73 dramas, 19 comedies, 8 tragedies, and 61 Chinese plus 39 international works. The genre counts (73+19+8=100) are consistent, but the paper should clarify how the same entries can be classified along multiple axes (e.g., musical vs. drama) and whether the retrieval/evaluation uses these labels.
  4. [§5.3] The text says 'We conducted parallel evaluations across all results using GPT-4o and Gemini-1.5-Pro, with detailed outcomes provided in the Appendix.' As a reviewer, I cannot verify this Appendix because it is not in the submitted manuscript. Please include those results in the supplementary material or main text; otherwise, the claim is unverifiable.

Circularity Check

1 steps flagged

VLM-as-a-Judge superiority is partially self-defined: Eq. (4) selects the output maximizing the same rubric used as the headline evaluation metric.

specific steps
  1. self definitional [§3.2 (Eq. 4), §5.1 evaluation metrics, Table 1]
    "Upon reaching the iteration limit K, the system evaluates the entire trajectory of generated images. It selects the poster with the highest cumulative evaluation score among the initial and all refined candidates. This decisive selection process is formalized as: I*_post = arg max ... Score(I). A_crit evaluates the image across three dimensions comprising eight sub-metrics (scored 1-10): Theatrical Aesthetics (abstraction, tension, refinement), Communication Efficacy (integration, clarity), and Thematic Fidelity (resonance, aptness, contextuality). ... 'we utilize a state-of-the-art VLM as an"

    Eq. (4) explicitly selects the poster maximizing Score(I), where Score is the Critic's VLM judgment on the same three dimensions and eight sub-metrics later used as the evaluation dimensions in §5.1 and Table 1. Thus the reported VLM-as-a-Judge superiority for Theatrical Aesthetics, Communication Efficacy, and Thematic Fidelity is not an independent measurement: the final output was chosen to maximize the same rubric (or a rubric with identical categories). Baselines are not given this argmax selection, so the comparison is biased by construction. The same Critic also drives background and portrait refinement toward functional whitespace and expressiveness, which are later scored as Whitespace and Expressiveness in Table 1. The human study (49 participants) and objective CLIP-I/DINOv2/OCR

full rationale

The central quantitative claim of superiority in semantic fidelity, aesthetic consistency, and interactivity is supported mainly by VLM-as-a-Judge metrics (Table 1). The derivation chain for those metrics is circular: ETPDesigner's Critic iteratively scores and optimizes generated posters on the exact three dimensions and eight sub-metrics (Theatrical Aesthetics, Communication Efficacy, Thematic Fidelity) that are later used as the evaluation dimensions, and Eq. (4) selects the candidate with the highest such score. Hence the method is explicitly tuned to the evaluation instrument; the baselines are not. The ablation study further confirms that removing the Critic lowers VLM scores, showing the optimization is chasing the evaluation metric. The only non-circular evidence is a small human study (49 participants) and objective consistency/text metrics (CLIP-I, DINOv2, OCR), which are not part of the VLM-judge loop and do lend some independent support. However, because the paper's headline 'superiority' in aesthetics and fidelity is measured almost entirely through the same VLM rubric the system optimizes, the evaluation is partially circular, warranting a score of 7 rather than a lower score. I do not count the RAG retrieval from ETP-Pro as circularity per se, though it is a benchmark-leakage concern.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The paper introduces no trained parameters, but it depends on several hand-chosen constants and domain assumptions. The most notable free parameters are the threshold 6.5, K=3, RAG top-k=3, and the hand-crafted expert rules. The invented 'global style anchor' is an internal mechanism that helps maintain consistency but has no independent evidence of capturing true style.

free parameters (4)
  • sub-metric pass threshold = 6.5
    In §3.2, the critic mandates optimization for any sub-metric scoring below 6.5; this threshold is hand-chosen, not learned or swept.
  • max optimization iterations K = 3
    In §3.2, the inner refinement loop runs at most K=3 times; K is a hand-chosen constant.
  • RAG retrieval top-k = 3
    In §3.2, the RAG module queries the three most similar posters; k=3 is a hand-chosen hyperparameter.
  • expert-system mapping rules = not specified numerically
    §3.1 defines Ψexpert as a deterministic mapping from emotion/genre/theme to color palette, composition, and style. The rules are crafted by hand and not learned; their exact definitions are not given.
axioms (4)
  • domain assumption VLM-as-a-Judge scores are valid proxies for human aesthetic quality in theatrical design
    §5.1 uses a VLM as the primary evaluator for aesthetics, fidelity, and expressiveness. The paper does not validate that VLM scores correlate with human ratings, despite the fact that the human study is only reported qualitatively.
  • domain assumption CLIP-I and DINOv2 similarities measure stylistic consistency across heterogeneous assets
    §5.1 uses CLIP-I and DINOv2 as objective metrics for global consistency and character-identity consistency. These metrics are known proxies, but their appropriateness for theater-program aesthetics is assumed.
  • domain assumption ETP-Pro is a representative sample of professional electronic theater programs
    §4 constructs ETP-Pro from 100 programs; no sampling protocol is given, and the benchmark itself is introduced by this paper, so representativeness is not independently established.
  • domain assumption LLM-based semantic extraction from raw scripts correctly identifies themes, emotions, and character traits
    §3.1 relies on ΦLLM to map script text into semantic representations without verification of extraction accuracy.
invented entities (1)
  • global style anchor vector s_global = (s_art, s_color) no independent evidence
    purpose: It is extracted from the final poster and used to condition all subsequent assets (backgrounds, character portraits) for stylistic consistency.
    This is a latent representation defined entirely within the system. It is not validated against external style annotations or shown to correspond to a human-interpretable style concept; its only evidence is the system's own downstream consistency scores.

pith-pipeline@v1.3.0-alltime-deepseek · 13631 in / 8825 out tokens · 80433 ms · 2026-08-01T11:11:22.955406+00:00 · methodology

0 comments
read the original abstract

Electronic Theater Programs (ETPs) serve as critical promotional media in the performing arts, comprising a multi-page collection of heterogeneous visual assets such as theatrical posters, performance details, and character portraits. However, existing text-to-image paradigms struggle with such complex design tasks due to their inability to comprehend long-context narratives and maintain visual consistency across multiple distinct pages. To address this, we introduce ETPDesigner, a collaborative Multi-Agent framework that directly synthesizes high-quality ETPs from raw dramatic scripts. Emulating a professional design pipeline, our framework orchestrates specialized agents for semantic script analysis, core poster synthesis, functional background generation, and the stratified composition of character assets. Central to ETPDesigner is a global style anchor mechanism that extracts visual priors from the core poster to enforce strict aesthetic uniformity across all generated components. Furthermore, we elevate the ETP from a static publication to an immersive interactive companion. By integrating portrait animation, customized speech synthesis, and persona-grounded Large Language Models (LLMs), our system enables users to engage in real-time, voice-enabled conversations with the generated virtual characters. To rigorously benchmark this task, we construct ETP-Pro, a domain-specific benchmark of professional theater posters and high-quality character portraits. Extensive evaluations demonstrate our method's superiority in producing semantically faithful, aesthetically consistent, and highly interactive program sets.

Figures

Figures reproduced from arXiv: 2607.19947 by Chaofeng Chen, Mengtian Li, Xiao Rong, Xiaoru Lin, Xinru Guo, Zhifeng Xie.

Figure 1
Figure 1. Figure 1: Overcoming the steep threshold of manual design, ETPDesigner automates ETP generation through multi-agent [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the ETPDesigner framework. ETPDesigner first extracts semantic profiles and design constraints from [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: ETP-Pro Benchmark. We curated a collection of ex [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison between ETPDesigner, Nano Banana Pro, Seedream 4.5, and PosterCraft. The visual results [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of Three Methods Across Four Dimen [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Applications:Transforming static images into char [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 9 linked inside Pith

  1. [1]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al . 2025. Qwen3-VL Technical Report.arXiv preprint arXiv:2511.21631(2025)

    Pith/arXiv arXiv 2025

  2. [2]

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. 2023. Instructpix2pix: Learning to follow image editing instructions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 18392–18402

    2023

  3. [3]

    Chieh-Yun Chen, Min Shi, Gong Zhang, and Humphrey Shi. 2025. T2i-copilot: A training-free multi-agent text-to-image system for enhanced prompt interpre- tation and interactive generation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 19396–19405

    2025

  4. [4]

    Haoyu Chen, Xiaojie Xu, Wenbo Li, Jingjing Ren, Tian Ye, Songhua Liu, Ying- Cong Chen, Lei Zhu, and Xinchao Wang. 2025. Posta: A go-to framework for customized artistic poster generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Conference (CVPR). 28694–28704

    2025

  5. [5]

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhong- dao Wang, James Kwok, Ping Luo, Huchuan Lu, et al . 2024. PixArt- 𝛼: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. InThe Thirteenth International Conference on Learning Representations (ICLR), Vol. 2024. 57611–57640

    2024

  6. [6]

    Minghao Chen, Iro Laina, and Andrea Vedaldi. 2024. Training-free layout control with cross-attention guidance. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). 5343–5353

    2024

  7. [7]

    SiXiang Chen, Jianyu Lai, Jialin Gao, Tian Ye, Haoyu Chen, Hengyu Shi, Shi- tong Shao, Yunlong Lin, Song Fei, Zhaohu Xing, et al . 2026. PosterCraft: Re- thinking High-Quality Aesthetic Poster Generation in a Unified Framework. InThe Fourteenth International Conference on Learning Representations (ICLR). https://openreview.net/forum?id=GhqnOEXQh3

    2026

  8. [8]

    Junhao Cheng, Xi Lu, Hanhui Li, Khun Loun Zai, Baiqiao Yin, Yuhao Cheng, Yiqiang Yan, and Xiaodan Liang. 2024. Autostudio: Crafting consistent subjects in multi-turn interactive image generation.arXiv preprint arXiv:2406.01388(2024)

    Pith/arXiv arXiv 2024

  9. [9]

    Yutao Cheng, Zhao Zhang, Maoke Yang, Hui Nie, Chunyuan Li, Xinglong Wu, and Jie Shao. 2025. Graphic design with large multimodal model. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), Vol. 39. 2473–2481

    2025

  10. [10]

    Yusuf Dalva, Guocheng Gordon Qian, Maya Goldenberg, Tsai-Shien Chen, Kfir Aberman, Sergey Tulyakov, Pinar Yanardag, and Kuan-Chieh Jackson Wang. 2025. Canvas-to-Image: Compositional Image Generation with Multimodal Controls. arXiv preprint arXiv:2511.21691(2025)

    arXiv 2025

  11. [11]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis.Advances in Neural Information Processing Systems (NeurIPS)34 (2021), 8780–8794

    2021

  12. [12]

    David Dinkevich, Matan Levy, Omri Avrahami, Dvir Samuel, and Dani Lischinski

  13. [13]

    Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. 2024. DreamLLM: Synergistic Multimodal Comprehension and Creation. InInternational Conference on Machine Learning (ICML), Vol. 2024. 6666–6702

    2024

  14. [14]

    Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. 2024. Instructdiffusion: A generalist modeling interface for vision tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 12709–12720

    2024

  15. [15]

    Google DeepMind. 2025. Nano Banana (Gemini 2.5 Flash Image): Google Image Editing and Generation Model. https://ai.google.dev/gemini-api/docs/image- generation. Accessed: 2026-07-17

    2025

  16. [16]

    Huiguo He, Huan Yang, Zixi Tuo, Yuan Zhou, Qiuyue Wang, Yuhang Zhang, Zeyu Liu, Wenhao Huang, Hongyang Chao, and Jian Yin. 2025. Dreamstory: Open-domain story visualization by llm-guided multi-subject consistent diffusion. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)(2025)

    2025

  17. [17]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems (NeurIPS)33 (2020), 6840–6851

    2020

  18. [18]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card.arXiv preprint arXiv:2410.21276(2024)

    Pith/arXiv arXiv 2024

  19. [19]

    Zeeshan Khan, Shizhe Chen, and Cordelia Schmid. 2025. ComposeAnything: Com- posite Object Priors for Text-to-Image Generation.arXiv preprint arXiv:2505.24086 (2025)

    Pith/arXiv arXiv 2025

  20. [20]

    Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. 2023. Generating images with multimodal language models.Advances in Neural Information Processing Systems (NeurIPS)36 (2023), 21487–21506

    2023

  21. [21]

    Black Forest Labs. 2024. FLUX. https://github.com/black-forest-labs/flux

    2024

  22. [22]

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. 2023. Gligen: Open-set grounded text-to-image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 22511–22521

    2023

  23. [23]

    Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. 2023. Autoposter: A highly automatic and content-aware design system for advertising poster generation. InProceedings of the 31st ACM International Conference on Multimedia (ACM MM). 1250–1260

    2023

  24. [24]

    Chang Liu, Haoning Wu, Yujie Zhong, Xiaoyun Zhang, Yanfeng Wang, and Weidi Xie. 2024. Intelligent grimm-open-ended visual storytelling via latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 6190–6200

    2024

  25. [25]

    Junle Liu, Peirong Zhang, Yuyi Zhang, Pengyu Yan, Hui Zhou, Xinyue Zhou, Fengjun Guo, and Lianwen Jin. 2026. PosterVerse: A Full-Workflow Framework for Commercial-Grade Poster Generation with HTML-Based Scalable Typography. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), Vol. 40. 7197–7205

    2026

  26. [26]

    Jian Ma, Yonglin Deng, Chen Chen, Nanyang Du, Haonan Lu, and Zhenyu Yang

  27. [27]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2024. Sdxl: Improving latent diffusion models for high-resolution image synthesis. InThe Twelfth International Conference on Learning Representations (ICLR), Vol. 2024. 1862–1874

    2024

  28. [28]

    InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), Vol

    Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), Vol. 39. 5955–5963

  29. [29]

    Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. 2024. Emu edit: Precise image editing via recognition and generation tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8871–8879

    2024

  30. [30]

    Team Seedream, Yunpeng Chen, Yu Gao, Lixue Gong, Meng Guo, Qiushan Guo, Zhiyao Guo, Xiaoxia Hou, Weilin Huang, Yixuan Huang, et al. 2025. Seedream 4.0: Toward Next-generation Multimodal Image Generation.arXiv preprint arXiv:2509.20427(2025)

    Pith/arXiv arXiv 2025

  31. [31]

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530(2024)

    Pith/arXiv arXiv 2024

  32. [32]

    Yucheng Suo, Fan Ma, Kaixin Shen, Linchao Zhu, and Yi Yang. 2025. Long- horizon Visual Instruction Generation with Logic and Attribute Self-reflection. InThe Thirteenth International Conference on Learning Representations (ICLR). https://openreview.net/forum?id=EdMb9TqqDY

    2025

  33. [33]

    Wen Wang, Canyu Zhao, Hao Chen, Zhekai Chen, Kecheng Zheng, and Chunhua Shen. 2025. Autostory: Generating diverse storytelling images with minimal human efforts.International Journal of Computer Vision (IJCV)133, 6 (2025), 3083–3104

    2025

  34. [34]

    Z-Image Team. 2025. Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer.arXiv preprint arXiv:2511.22699(2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. 2024. Genartist: Multimodal llm as an agent for unified image generation and editing.Advances in Neural Information Processing Systems (NeurIPS)37 (2024), 128374–128395

    2024

  36. [36]

    Zhendong Wang, Jianmin Bao, Shuyang Gu, Dong Chen, Wengang Zhou, and Houqiang Li. 2025. DesignDiffusion: High-Quality Text-to-Design Image Gen- eration with Diffusion Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 20906–20915

    2025

  37. [37]

    Jianzong Wu, Chao Tang, Jingbo Wang, Yanhong Zeng, Xiangtai Li, and Yunhai Tong. 2025. Diffsensei: Bridging multi-modal llms and diffusion models for customized manga generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Conference (CVPR). 28684–28693

    2025

  38. [38]

    Jiazhe Wei, Ken Li, Tianyu Lao, Haofan Wang, Liang Wang, Caifeng Shan, and Chenyang Si. 2025. PosterCopilot: Toward Layout Reasoning and Controllable Editing for Professional Graphic Design.arXiv preprint arXiv:2512.04082(2025)

    arXiv 2025

  39. [39]

    Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. 2024. Self-correcting llm-controlled diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 6327–6336

    2024

  40. [40]

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al . 2023. AutoGen: En- abling Next-Gen LLM Applications via Multi-Agent Conversation.arXiv preprint arXiv:2308.08155(2023)

    Pith/arXiv arXiv 2023

  41. [41]

    Meng Xu, Tong Zhang, Fuyun Wang, Yi Lei, Xin Liu, and Zhen Cui. 2025. MPDS: A Movie Posters Dataset for Image Generation with Diffusion Model. InChinese Conference on Pattern Recognition and Computer Vision (PRCV). Springer, 138–151

    2025

  42. [42]

    Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. 2023. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV). 7452–7461. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Li et al

    2023

  43. [43]

    Shuai Yang, Yuying Ge, Yang Li, Yukang Chen, Yixiao Ge, Ying Shan, and Ying- Cong Chen. 2025. Seed-story: Multimodal long story generation with large language model. InProceedings of the IEEE/CVF International Conference on Com- puter Vision (ICCV). 1850–1860

    2025

  44. [44]

    Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui

  45. [45]

    Zhengyuan Yang, Jianfeng Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. 2024. Idea2img: Iterative self-refinement with gpt-4v for automatic image design and generation. InEuropean Conference on Computer Vision (ECCV). Springer, 167–184

    2024

  46. [46]

    Hui Zhang, Dexiang Hong, Maoke Yang, Yutao Cheng, Zhao Zhang, Jie Shao, Xinglong Wu, Zuxuan Wu, and Yu-Gang Jiang. 2025. Creatidesign: A unified multi-conditional diffusion transformer for creative graphic design.arXiv preprint arXiv:2505.19114(2025)

    arXiv 2025

  47. [47]

    Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. 2023. Glyphcontrol: Glyph conditional control for visual text generation.Advances in Neural Information Processing Systems (NeurIPS)36 (2023), 44050–44066

    2023

  48. [48]

    Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou

  49. [50]

    Yiming Zhao, Yuanpeng Gao, Yuxuan Luo, Jiwei Duan, Shisong Lin, Longfei Xiong, and Zhouhui Lian. 2025. UTDesign: A Unified Framework for Stylized Text Editing and Generation in Graphic Design Images. InProceedings of the SIGGRAPH Asia 2025 Conference Papers (SIGGRAPH Asia). 1–11

    2025

  50. [52]

    Storydiffusion: Consistent self-attention for long-range image and video generation.Advances in Neural Information Processing Systems (NeurIPS)37 (2024), 110315–110340

    2024

  51. [2024]

    InInternational Conference on Machine Learning (ICML), Vol

    Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. InInternational Conference on Machine Learning (ICML), Vol. 3. 7

  52. [2025]

    arXiv preprint arXiv:2508.09983(2025)

    Story2board: A training-free approach for expressive storyboard generation. arXiv preprint arXiv:2508.09983(2025)

    Pith/arXiv arXiv 2025