PosterCraft: Rethinking High-Quality Aesthetic Poster Generation in a Unified Framework

Haoyu Chen; Hengyu Shi; Jialin Gao; Jianyu Lai; Junfeng Luo; Lei Zhu; Shitong Shao; Sixiang Chen; Song Fei; Tian Ye

REVIEW 3 major objections 5 minor 3 cited by

PosterCraft: Rethinking High-Quality Aesthetic Poster Generation in a Unified Framework

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read PosterCraft claims that a single diffusion model, trained through four stages, can generate finished aesthetic posters—typography, artistic content, and layout—directly from one prompt.

desk verdict A well-engineered poster-generation recipe whose headline numbers rest on a Gemini OCR protocol that primes the evaluator with the expected text — worth engaging, but the Table 1 claims need independent verification. read the letter →

arxiv 2506.10741 v1 pith:P6RAJNFN submitted 2025-06-12 cs.CV

SiXiang Chen , Jianyu Lai , Jialin Gao , Tian Ye , Haoyu Chen , Hengyu Shi , Shitong Shao , Yunlong Lin

show 6 more authors

Song Fei Zhaohu Xing Yeying Jin Junfeng Luo Xiaoming Wei Lei Zhu

This is my paper · ORCID

classification cs.CV

keywords postergenerationtextrenderingdiffusionmodelsdirectpreferenceoptimizationvision-languagefeedbackregion-awarecalibrationflowmatchingautomateddatasetconstruction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

PosterCraft sets out to show that high-quality aesthetic posters do not need modular pipelines, layout planners, or predefined templates: a single diffusion model can generate finished posters, including precise typography and artistic layout, directly from a prompt. The paper builds a four-stage cascaded workflow on the Flux.1-dev backbone—text-rendering fine-tuning on two million synthetic samples, region-aware supervised fine-tuning, best-of-n preference optimization, and vision-language feedback refinement—with each stage supported by an automated dataset construction pipeline. On a 300-image Gemini-generated benchmark, PosterCraft reports text recall 0.787, text F-score 0.774, and text accuracy 0.735, above all tested open-source systems and close to Gemini2.0-Flash-Gen (0.798, 0.786, 0.746). The claim matters because it suggests that strong text rendering and aesthetic coherence can be achieved in an end-to-end generative pass rather than by composing separately generated backgrounds and text overlays.

What carries the argument

The load-bearing mechanism is the four-stage cascade itself, applied to the Flux.1-dev diffusion backbone without architectural changes. Stage 1 fine-tunes the model on Text-Render-2M with the flow-matching loss to make text rendering reliable. Stage 2 applies region-aware calibration, which multiplies the flow-matching loss by a per-pixel weight map $w(p)$ that is $0.6$ inside large text masks, $0.2$ inside small text masks, and $1.0$ everywhere else; this is the named identity that balances text fidelity against stylistic coherence. Stage 3 optimizes the DPO objective over best-of-5 and worst-of-5 poster pairs selected by HPSv2 and filtered by a text-accuracy check, injecting global aesthetic preferences that per-pixel losses cannot express. Stage 4 conditions generation on a joint vision-language context $c=[e_p; e_{c,s}; v_{\text{img}}]$, concatenating the prompt embedding with T5-encoded content and style reflections and a VAE-encoded image feedback signal, and fine-tunes under the conditional flow-matching loss. The cascade is what converts a general text-to-image model into a poster-specific generator.

What would settle it

Have several human transcribers independently transcribe every text string on the 300 test posters and recompute recall, F-score, and accuracy from those transcripts; if the human scores diverge materially from the Gemini-based scores on stylized fonts, the quantitative claim is evaluator-dependent.

Watch

Extended reading notes

Core claim

PosterCraft's central claim is that a unified, cascaded optimization workflow can unlock the poster-generation ability of a standard diffusion backbone without architectural modification. The paper asserts that each of its four stages addresses a distinct bottleneck: flow-matching fine-tuning on Text-Render-2M supplies reliable text rendering; region-aware calibration with a per-pixel weight map (0.6 for large text, 0.2 for small text, 1.0 elsewhere) preserves text fidelity while letting non-text regions define the poster's aesthetic; aesthetic-text reinforcement learning via DPO on HPSv2-scored best-of-5 and worst-of-5 pairs injects higher-order preferences for layout balance, color harmony, and typographic cohesion; and joint vision-language conditioning, which encodes Gemini-generated content and style reflections through T5 and injects VAE-encoded image feedback, enables iterative refinement. The paper reports that this workflow yields text recall 0.787, F-score 0.774, and accuracy 0.735 on 300 test images, outperforming open-source baselines and approaching the closed-source Gemini2.0-Flash-Gen.

Load-bearing premise

All reported text metrics assume Gemini's OCR reads artistic poster typography accurately enough to serve as ground truth, so a systematic OCR blind spot would change the headline numbers.

Editorial extensions

If this is right

End-to-end poster generation becomes practical: a user supplies one prompt, and the model decides text placement, typographic style, and layout without an external layout planner or predefined templates.
The same automated dataset construction—synthetic text-render data, MLLM-filtered poster corpora, HPSv2-scored preference pairs, and Gemini-reflected feedback—can be rebuilt for other text-heavy visual domains, since the paper stresses the pipelines are fully automated.
The reported text metrics imply open-source poster models can approach a commercial system's typography, narrowing the gap that previously required closed APIs.
The vision-language feedback loop at inference allows a generated poster to be critiqued and regenerated, giving designers an iterative refinement mechanism rather than a one-shot output.
The workflow is explicitly transferable to stronger backbones, so future improvements to base diffusion models should compound with the four stages.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper evaluates text with Gemini's own OCR; an independent character-level transcriber on the same 300 images would show whether the reported gains are evaluator-specific, a test the paper does not run.
The hand-set region weights (0.6 for large text, 0.2 for small text, 1.0 elsewhere) are a single operating point; the paper reports no sensitivity sweep, so it is unknown how robust the text-aesthetic trade-off is to nearby weight choices.
The best-of-n preference construction selects pairs by HPSv2 score filtered by Gemini text checks; if human aesthetic judgment diverges from HPSv2 on artistic poster styles, the DPO signal could encode a biased notion of quality.
The same cascade could plausibly apply to book covers, UI mockups, or signage, where text and layout harmony matter as much as in posters; the paper does not test these adjacent tasks.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: PosterCraft is a well-engineered recipe with a real chance of being useful, but the headline numbers should be treated with caution because the OCR evaluator is told what text to expect before it reads the image. The four-stage cascade — synthetic text overlay fine-tuning, region-aware SFT, DPO with HPSv2, and VLM reflection — is a coherent and genuinely new combination, and the datasets are large and documented. The ablations show each stage contributes, and the visual comparisons are striking.

The soft spot is evaluation. Prompt 11.1 in the supplementary extracts the ground-truth strings from the design prompt, then asks Gemini to OCR the image in the same context. That is a textbook confirmation-bias setup: a multimodal model primed with the expected title will often 'see' it even when glyphs are distorted or missing. Because the same commercial family generates the test prompts (Gemini2.0-Flash-Gen), serves as the top competitor, and supplies the OCR and preference evaluator (Gemini2.5-Flash), the Table 1 numbers are at least partly self-referential. The gap to Gemini (0.787 vs 0.798 recall) is small enough that evaluator bias could flip the ranking. No independent OCR, no human character-level transcription, no confidence intervals. The user study of 20 designers is mentioned but not reported with counts, so we can't weigh it properly.

That said, the paper isn't a sham. The recipe is specific enough to reproduce once they release code, models, and datasets as promised. The region-aware weighting and the HPSv2-based preference pair construction are reasonable, but a few thresholds (0.6/0.2/1.0, HPSv2 0.25, score-diff 0.025) are asserted without sensitivity analysis. Those are minor compared to the evaluation issue.

Bottom line: this deserves a serious referee, but the referee should demand independent OCR (or human transcription), per-pair counts from the user study, and the release of artifacts. If the numbers hold under independent measurement, it's a solid contribution for anyone building on open diffusion backbones. My recommendation: send to peer review with 'revise' — the core method is strong but the empirical foundation needs to be made independent of the vendor being compared.

Referee Report

3 major / 5 minor

Summary. The paper proposes PosterCraft, a unified end-to-end pipeline for aesthetic poster generation built on the Flux.1-dev diffusion backbone. The method consists of four cascaded stages: (i) large-scale text-rendering fine-tuning on a new Text-Render-2M dataset, (ii) region-aware supervised fine-tuning on HQ-Poster-100K with per-pixel loss weights for large text, small text, and background, (iii) aesthetic-text reinforcement learning via best-of-n DPO on Poster-Preference-100K, and (iv) joint vision-language feedback refinement using Poster-Reflect-120K and a fine-tuned InternVL3-8B. The authors claim state-of-the-art performance among open-source systems and near-commercial quality, reporting text recall 0.787, text F-score 0.774, and text accuracy 0.735 on a 300-image Gemini-generated test set, approaching Gemini2.0-Flash-Gen (0.798/0.786/0.746). The paper also reports a user study, a Gemini-based preference evaluation, qualitative comparisons, and ablations, and states that code, models, and datasets will be released.

Significance. If the reported results hold, PosterCraft is a substantial contribution: it demonstrates that a standard diffusion backbone can be trained end-to-end for complex poster generation without modular layout planning, and the four-stage workflow together with the fully automated dataset-construction pipelines could serve as a practical template for design-oriented generation. The release of code, models, and datasets would be valuable to the community. The paper is also careful to include ablation studies for each stage and detailed supplementary prompts. However, the central quantitative claims rest on an evaluation protocol that is vulnerable to confirmation bias, and the evaluation loop uses the same commercial model family for prompt generation, OCR scoring, and preference evaluation. These issues undermine the current evidence for the headline numbers, even though the qualitative results and ablations provide partial support. The contribution is potentially significant, but the quantitative claims need to be verified with independent and unbiased measurement.

major comments (3)

[Section 4.2 and Supplementary Prompt 11.1] The Table 1 headline metrics rely on an OCR protocol in which Gemini2.5-Flash is first given the full design prompt and instructed to extract the exact ground-truth strings, and then, in the same context, is asked to run OCR on the generated image and perform character-level alignment. This primes the evaluator with the expected text and is highly susceptible to confirmation bias: the model may read distorted, missing, or stylized glyphs as the expected characters, and the bias need not affect all models equally. Since the same model family (Gemini) also generated the test prompts and is the leading closed-source comparison target, the reported differences in Table 1 may be artifacts of the measurement protocol rather than genuine text-rendering gains. The authors should repeat the evaluation with an independent OCR engine (e.g., PaddleOCR or Tesseract) and/or human character-level transcription on a random subset, and report per-model confidence intervals for recall, F-score, and accuracy.
[Section 4.2 and Figure 6] The Gemini-based preference evaluation (Fig. 6) uses Gemini2.5-Flash as an 'authoritative evaluator' without any validation of its agreement with the human designers who participated in the user study (Fig. 5). Because Gemini2.5-Flash belongs to the same model family as Gemini2.0-Flash-Gen, the win rates in Fig. 6 could be biased toward outputs that match Gemini's aesthetic and text-rendering preferences rather than reflecting true human preferences. The authors should report human-machine agreement (e.g., Cohen's kappa on a shared subset of comparisons) and provide exact win/tie/loss counts with binomial confidence intervals for both the user study and the Gemini evaluation.
[Section 4.2] The entire quantitative comparison is based on 300 test images (100 prompts, 3 samples per model), and no confidence intervals, standard deviations, or per-prompt variability are reported. Given the small sample size and the fact that the prompts are themselves generated by Gemini2.0-Flash-Gen, the 0.01-0.02 gaps between PosterCraft and Gemini2.0-Flash-Gen in Table 1 are not statistically established. The authors should report bootstrap or other confidence intervals for the three text metrics, and ideally include a human-curated set of poster prompts from diverse real-world sources to test generalization beyond synthetic Gemini-generated prompts.

minor comments (5)

[Section 3.2, Eq. (3)] The region-aware calibration weights 0.6, 0.2, and 1.0 are central to the stage-2 loss, but no sensitivity analysis is provided; please report at least one ablation that varies these weights to demonstrate that the chosen values are not brittle.
[Section 4.1] The Magic Prompter is described only as 'Qwen3-based'; please specify the exact model and version used at inference, since it is part of the generation pipeline and could affect reproducibility.
[Section 4.2 and Supplementary Section 11] The main text says the OCR engine is 'the SOTA VLM [30]' while the supplementary specifies Gemini2.5-Flash-Preview-05-20; please disambiguate the exact model and version used for the reported metrics.
[Figures 5 and 6] The win-rate plots in Figs. 5 and 6 are difficult to read at print resolution; consider providing the exact numerical win/tie/loss counts in a table in the supplementary material.
[Section 13 (Limitations)] The limitations section is brief and does not mention the reliance on Gemini-generated data and Gemini-based evaluation; please add a candid discussion of these dependencies and their potential impact on the reported results.

Circularity Check

2 steps flagged · score 6.0 of 10

Table 1's headline text metrics are computed by a Gemini OCR evaluator that is primed with the expected strings in the same prompt, while the test prompts and the leading closed-source competitor are also Gemini; the quantitative claim is partially circular.

other [Supplementary Sec. 11, Prompt 11.1 (OCR Evaluation); used for Sec. 4.2 / Table 1]
"(1) Ground-Truth Extraction (from the design prompt only): •Do NOT read text from the image for GT. • Parse ONLY the following design prompt and extract ALL text strings that should appear on the poster (titles, subtitles, dates, slogans, venue, etc.), preserving spaces and punctuation exactly: `original_prompt_text` ... (2) OCR Extraction (from the attached image): •Run OCR on the provided image and extract ALL rendered text exactly as it appears."

The OCR output that feeds Table 1 is produced in the same Gemini2.5-Flash context that has just been told the exact strings expected on the poster. Because an autoregressive MLLM conditions on the preceding prompt, the OCR_text (and therefore C, D, S, and the reported recall/F/accuracy) can be pulled toward the expected GT even when glyphs are distorted, missing, or stylized. The claimed numbers—PosterCraft 0.787 recall/0.774 F/0.735 accuracy versus Flux1.dev 0.723/0.707/0.667 and Gemini2.0-Flash-Gen 0.798/0.786/0.746—are thus not independent measurements but partly a function of the evaluator's priming. No independent OCR engine or human character-level transcription is reported, so the central quantitative claim reduces, in part, to the evaluator's own prior context.
other [Sec. 4.2 (test-prompt generation and baselines), Table 1, Supplementary Sec. 11]
"To assemble our test prompts, we use Gemini2.0-Flash-Gen [30] to randomly generate 100 aesthetic poster prompts—balanced across short, medium, and long lengths—and for each prompt we sample three output per model, yielding 300 test images. We generate posters with OpenCOLE [11], Playground-v2.5 [17], SD3.5 [1], Flux1.dev [9], Ideogram-v2 [32], BAGEL [5], and Gemini2.0-Flash-Gen [30], apply the OCR engine of the SOTA VLM [30] to each image ... we employ Gemini2.5-Flash-Preview-05-20 as both an OCR metric calculator and a multi-dimensional preference evaluator."

The evaluation loop is self-referential at the system level: the same Gemini family writes the test prompts, provides the strongest closed-source baseline, and (via Gemini2.5-Flash) computes the OCR metrics used in Table 1. This means the comparison against Gemini2.0-Flash-Gen is not an external benchmark but a within-family measurement, and the test distribution may be matched to Gemini's output style. In addition, Gemini2.5-Flash is used throughout data construction (mask generation, prompt-alignment verification, reflection feedback), so the final evaluation metric is produced by the same model family that shaped the training data. This makes the headline 'approaching Gemini2.0-Flash-Gen' partially an artifact of the evaluation chain rather than an independent first-principles result.

full rationale

The four training stages themselves are not circular: Text-Render-2M is synthetic with text rendered by construction, region-aware SFT is a standard weighted flow-matching loss, DPO uses HPSv2 plus Gemini verification, and reflection fine-tuning uses Gemini feedback—these are legitimate training signals. The self-citation to POSTA [2] appears only in related-work context and is not load-bearing; there is no uniqueness theorem or ansatz-by-citation. The circularity is concentrated in the evaluation protocol: the single most important quantitative claim (Table 1) is produced by an OCR evaluator that is primed with the expected text in the same prompt and that belongs to the same model family as the prompt generator and leading competitor. The paper's own Limitations section acknowledges reliance on the Flux backbone but does not acknowledge this evaluator loop. Because the user study and qualitative comparisons provide some independent support, the paper is not wholly circular, but the headline numeric superiority is partially constructed by the evaluation setup.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are invented. The paper's contributions are datasets and a training pipeline. The free parameters are the numerous hand-chosen thresholds and weights that control dataset filtering and region-aware training. The axioms are the reliability of the automated judges (Gemini and HPSv2) and the sufficiency of the FLUX backbone.

free parameters (6)

Region-aware calibration weights = 0.6 (large text), 0.2 (small text), 1.0 (non-text)
The per-pixel weights in Eq. (3) are chosen by hand. No sweep, sensitivity analysis, or justification is provided for these exact values. They directly shape the trained model and the reported ablation relies on the choice.
HPSv2 threshold 0.25 for HQ-Poster-100K filtering = 0.25
Posters scoring below 0.25 in HPS are filtered out. The threshold is selected by hand with no analysis of how sensitive the final model is to it.
MLLM scorer threshold 0.98 for billing/credit block filtering = 0.98
Softmax probability threshold chosen by hand to control filtering stringency; its effect on the resulting dataset is not evaluated.
HPSv2 score difference threshold 0.025 for preference pair qualification = 0.025
Preference pairs are kept only if the HPSv2 score difference exceeds 0.025. The threshold is chosen by hand with no sensitivity analysis. It directly controls the DPO training data.
DPO temperature beta = not stated
The DPO objective in Eq. (6) includes a scaling factor beta that is not specified in the paper. The trained model depends on its value.
LoRA ranks = 64 (RL), 128 (VLM feedback)
LoRA ranks are architecture choices rather than fitted to data, but they are tuned for the pipeline and no ablation over ranks is shown.

assumptions (5)

domain assumption FLUX.1-dev is a capable base model for poster generation after fine-tuning; its prior text rendering and aesthetic failures are fixable with the proposed stages
The entire method is built on Flux-dev and the conclusions are conditional on the quality of this backbone. The paper explicitly notes this in Sec. 13 (Limitations).
domain assumption Gemini2.5-Flash provides accurate OCR text extraction and metric computation for artistic poster fonts
All numeric text metrics in Table 1 come from Gemini2.5-Flash's own OCR and metric calculation described in Sec. 11 and Prompt 11.1. If this OCR is inaccurate on artistic fonts, all text metrics are unreliable.
domain assumption Gemini2.5-Flash and HPSv2 preference scores reflect human aesthetic judgment
Preference pairs, best-of-6 selections, reflection targets, and final evaluations all rely on these automated judges. The paper provides a small human study for final comparison, but the training signals and main metrics are automated.
standard math DPO with best-of-n sampled pairs from the current diffusion policy improves generation quality
The paper follows prior diffusion-DPO work, using the ELBO over the diffusion chain. This is a published technique, but it is an unproved background assumption for this setup.
domain assumption The test prompts from Gemini2.0-Flash-Gen represent the distribution of real poster design requests
The benchmark is 100 prompts sampled by Gemini2.0-Flash-Gen, balanced by length. There is no external validation that these prompts are representative of actual user posters.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PosterCraft: Rethinking High-Quality Aesthetic Poster Generation in a Unified Framework." pith.science (2026). https://pith.science/paper/P6RAJNFN

@misc{pith2026250610741,
  author       = {Pith},
  title        = {Pith review of: PosterCraft: Rethinking High-Quality Aesthetic Poster Generation in a Unified Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P6RAJNFN}},
  note         = {Machine review of arXiv:2506.10741}
}

read the original abstract

Generating aesthetic posters is more challenging than simple design images: it requires not only precise text rendering but also the seamless integration of abstract artistic content, striking layouts, and overall stylistic harmony. To address this, we propose PosterCraft, a unified framework that abandons prior modular pipelines and rigid, predefined layouts, allowing the model to freely explore coherent, visually compelling compositions. PosterCraft employs a carefully designed, cascaded workflow to optimize the generation of high-aesthetic posters: (i) large-scale text-rendering optimization on our newly introduced Text-Render-2M dataset; (ii) region-aware supervised fine-tuning on HQ-Poster100K; (iii) aesthetic-text-reinforcement learning via best-of-n preference optimization; and (iv) joint vision-language feedback refinement. Each stage is supported by a fully automated data-construction pipeline tailored to its specific needs, enabling robust training without complex architectural modifications. Evaluated on multiple experiments, PosterCraft significantly outperforms open-source baselines in rendering accuracy, layout coherence, and overall visual appeal-approaching the quality of SOTA commercial systems. Our code, models, and datasets can be found in the Project page: https://ephemeral182.github.io/PosterCraft

Figures

Figures reproduced from arXiv: 2506.10741 by the authors.

**Figure 1.** Aesthetic posters generated by PosterCraft demonstrate that backgrounds, layouts, and typographic designs are produced directly from textual input [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** The four datasets of PosterCraft across its four stages: (1) Text-Render-2M for text rendering optimization in the initial phase, (2) HQ-Poster-100K, comprising over 100K high-quality posters with masks and captions, (3) Poster-Preference-100K, yielding 6K high-quality preference pairs from 100K generated samples, and (4) Poster-Reflect-120K, constructing 64K feedback pairs from 120K generated posters. we unlock the… view at source ↗

**Figure 3.** Comparison of text rendering on poster typography, plain-text scenes, and long-form text posters. Each pair shows the Flux.1 dev baseline (left), exhibiting missing, repeated, or error text, alongside the optimized output (right) after our scalable text rendering optimization, demonstrating marked gains in text fidelity, alignment, and accuracy. then fine-tune foundation models on paired Text-Render-2M using the flo… view at source ↗

Figures from the paper (24 more)

**Figure 4.** Figure 4: The pipeline of PosterCraft, which has four stages: (1) Text Rendering Optimization to improve text accuracy and fidelity; (2) High-Quality Poster Fine-Tuning with region-aware calibration to poster styling across text and non-text regions; (3) Aesthetic-Text Reinforce…

**Figure 5.** Figure 5: User study comparisons between PosterCraft and both SOTA opensource and closed-source models. PosterCraft consistently outperforms all open-source baselines and several proprietary systems cross multiple dimensions, with performance marginally below that of the leadin…

**Figure 6.** Figure 6: Gemini serves as an authoritative evaluator for human preference comparisons across our method and other baselines. PosterCraft outperforms most state-of-the-art generative models in aesthetic coherence, prompt alignment, text rendering, and overall preference. It achi…

**Figure 7.** Figure 7: Visual comparison of different model outputs. Red boxes highlight misspelled or distorted text, while yellow boxes indicate redundant or missing text elements. Within the prompts, orange text denotes content and style requirements, and red text indicates textual elemen…

**Figure 8.** Figure 8: Visual comparison of different model outputs. Red boxes highlight misspelled or distorted text, while yellow boxes indicate redundant or missing text elements. Within the prompts, orange text denotes content and style requirements, and red text indicates textual elemen…

**Figure 9.** Figure 9: Visual comparison of different model outputs. Red boxes highlight misspelled or distorted text, while yellow boxes indicate redundant or missing text elements. Within the prompts, orange text denotes content and style requirements, and red text indicates textual elemen…

**Figure 10.** Figure 10: Aesthetic-text reinforcement learning (top row) and vision-language feedback (bottom row) qualitative comparisons. orange boxes denote the text biases. Different color fonts represent key feedback information from VLM. The top examples demonstrate that our reinforceme…

**Figure 11.** Figure 11: Ablation experiments on the core components of our workflow: removing Text Rendering Optimization, Region-aware Calibration, Aesthetic-Text RL, or Reflection leads to sustained declines in OCR accuracy (purple) and human preference (peach), demonstrating the effectiv…

**Figure 12.** Figure 12: More samples are shown, which are high-quality paired samples in Text-Render-2M. [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]

**Figure 13.** Figure 13: Prompt for MLLM Scorer in HQ-Poster-100K. Prompt 8.2 (Gemini Caption Generation) Please write a structured and detailed caption in a single paragraph for this poster, covering the following five aspects in order: Poster Content—Describe what is visually depicted. Post…

**Figure 14.** Figure 14: Prompt for Gemini Caption Generation in HQ-Poster-100K. is shown in Fig.17 Our criteria require that text in the Best of 5 posters must be completely accurate while maximizing alignment with both the content and aesthetic style requirements specified in the prompt. We…

**Figure 15.** Figure 15: More pictures are shown, which are high-quality paired samples with masks in HQ-Poster-100K. [PITH_FULL_IMAGE:figures/full_fig_p014_15.png]

**Figure 23.** Figure 23: Fig.23 [PITH_FULL_IMAGE:figures/full_fig_p014_23.png]

**Figure 16.** Figure 16: Prompt for Gemini Mask Generation in HQ-Poster-100K. be able to fully correct these shortcomings. That said, our workflow is highly unified and readily transferable to stronger baselines, ensuring full compatibility with other models in the community. 14 Future work …

**Figure 17.** Figure 17: Prompt for Prompt Alignment Evaluation in Poster-Preference-100K Poster-Preference-100K "This poster for *The Hunger Games* features a close-up of Katniss Everdeen, portrayed by Jennifer Lawrence, aiming a bow and arrow in a sun-drenched forest … The style is cinemati…

**Figure 18.** Figure 18: Additional Preference Pairs in Poster-Preference-100K. The images on the left are Rejected Samples, while those on the right are Preferred Samples. Orange text indicates textual content, and red text corresponds to content, style, or layout requirements [PITH_FULL_IM…

**Figure 19.** Figure 19: Prompt for Best-of-6 Selection in Poster-Reflect-120K [PITH_FULL_IMAGE:figures/full_fig_p017_19.png]

**Figure 20.** Figure 20: Prompt for Feedback Collection in Poster-Reflect-120K [PITH_FULL_IMAGE:figures/full_fig_p018_20.png]

**Figure 21.** Figure 21: Prompt for OCR Evaluation [PITH_FULL_IMAGE:figures/full_fig_p019_21.png]

**Figure 22.** Figure 22: Prompt for Preference Evaluation [PITH_FULL_IMAGE:figures/full_fig_p020_22.png]

**Figure 23.** Figure 23: Prompt for Preference Evaluation [PITH_FULL_IMAGE:figures/full_fig_p021_23.png]

**Figure 24.** Figure 24: Prompt for Final Retouching Instructions [PITH_FULL_IMAGE:figures/full_fig_p022_24.png]

**Figure 25.** Figure 25: Examples generated by our PosterCraft demonstrating high diversity and aesthetic quality across themes including education, entertainment, and science fiction. All generation results showcase genre-specific fidelity, text rendering, and layout aesthetic [PITH_FULL_IM…

**Figure 26.** Figure 26: Examples generated by our PosterCraft demonstrating high diversity and aesthetic quality across themes including movies, product, and virtual reality. All generation results showcase genre-specific fidelity, text rendering, and layout aesthetic. , [PITH_FULL_IMAGE:fi…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

FLUX-Reason-6M & PRISM-Bench: A Million-Scale Text-to-Image Reasoning Dataset and Comprehensive Benchmark
cs.CV 2025-09 conditional novelty 7.0 of 10

The authors build a 6M-image, 20M-caption reasoning dataset with generation chain-of-thought and a 7-track VLM-judged benchmark, then rank 19 text-to-image models.
Rethinking Layered Graphic Design Generation with a Top-Down Approach
cs.CV 2025-07 conditional novelty 6.0 of 10

Accordion decomposes AI-generated raster designs into editable background, object, and vectorized text layers using a VLM-driven top-down planning pipeline.
JarvisHub: An Open Harness for Canvas-Native Multimodal Creative Agents
cs.CV 2026-07 conditional novelty 5.0 of 10

JarvisHub open-sources a three-layer canvas-state, protocol-bridge, and agent-runtime harness so multimodal creative agents can inspect and update a shared editable project graph over long workflows.

Reference graph

Works this paper leans on

48 extracted references · 15 canonical work pages · cited by 3 Pith papers

[1]

Stability AI. 2024. Stable Diffusion 3.5 Large. https://huggingface.co/stabilityai/ stable-diffusion-3.5-large

work page 2024
[2]

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.arXiv preprint arXiv:2503.14908(2025)

arXiv 2025
[4]

Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei
[5]

Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan. 2025. Emerging Properties in Unified Multimodal Pretraining. arXiv:2505.14683 [cs.CV] https://arxiv.org/abs/2505.14683

arXiv 2025
[6]

Textdiffuser: Diffusion models as text painters.Advances in Neural Infor- mation Processing Systems36 (2023), 9353–9387

2023
[7]

Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. 2023. Layoutgpt: Compo- sitional visual planning and generation with large language models.Advances in Neural Information Processing Systems36 (2023), 18225–18250

work page 2023
[8]

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. 2024. PosterCraft: Rethinking High-Quality Aesthetic Poster Generation in a Unified Framework•11 Scaling rectified flow transformers for high-resolution image synthesis. InForty- first International Confe...

work page 2024
[9]

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

work page 2024

Show all 48 references

[10]

Yifan Gao, Zihang Lin, Chuanbin Liu, Min Zhou, Tiezheng Ge, Bo Zheng, and Hongtao Xie. 2025. PosterMaker: Towards High-Quality Product Poster Genera- tion with Accurate Text Rendering.arXiv preprint arXiv:2504.06632(2025)

2025 arXiv
[11]

Naoto Inoue, Kento Masui, Wataru Shimoda, and Kota Yamaguchi. 2024. Open- COLE: Towards Reproducible Automatic Graphic Design Generation.arXiv preprint arXiv:2406.08232(2024)

2024 arXiv
[12]

Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi
[13]

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Layoutdm: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10167–10176
[14]

Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Arsh Koneru, Yusuke Kato, Kazuki Kozuka, and Aditya Grover. 2025. Reflect-DiT: Inference-Time Scaling for Text-to-Image Diffusion Transformers via In-Context Reflection.arXiv preprint arXiv:2503.12271(2025)

2025 arXiv
[15]

InternVL Team. 2024. InternVL2.5: Expanding Performance Boundaries of Open- Source Multimodal Models with Model, Data, and Test-Time Scaling. https: //internvl.github.io/blog/2024-12-05-InternVL-2.5

2024
[16]

InternVL Team. 2025. InternVL3: Advancing Open-Source Multimodal Models with Native Multimodal Pretraining. https://internvl.github.io/blog/2025-04-11- InternVL-3.0

2025
[17]

Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. 2024. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695(2024)

2024 arXiv
[18]

Shuangqi Li, Hieu Le, Jingyi Xu, and Mathieu Salzmann. 2024. Enhancing Com- positional Text-to-Image Generation with Reliable Random Seeds.arXiv preprint arXiv:2411.18810(2024)

2024 arXiv
[19]

Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, Jian-Guang Lou, and Dongmei Zhang. 2023. Layoutprompter: awaken the design ability of large language models. Advances in Neural Information Processing Systems36 (2023), 43852–43879

2023
[20]

Chuwei Luo, Yufan Shen, Zhaoqing Zhu, Qi Zheng, Zhi Yu, and Cong Yao. 2024. LayoutLLM: Layout Instruction Tuning with Large Language Models for Doc- ument Understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15630–15640

2024
[21]

Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. 2024. Glyph-ByT5-v2: A Strong Aesthetic Baseline for Accurate Multilingual Visual Text Rendering.arXiv preprint arXiv:2406.10208(2024)

2024 arXiv
[22]

Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

2017 arXiv
[23]

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems36 (2023), 53728–53741

2023
[24]

Yiyang Ma, Xingchao Liu, Xiaokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Liang Zhao, et al. 2024. Janusflow: Har- monizing autoregression and rectified flow for unified multimodal understanding and generation.arXiv preprint arXiv:2411.07975(2024)

2024 arXiv
[25]

Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952 (2023)

2023 arXiv
[26]

Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. InInternational Conference on Machine Learning. PMLR, 4596–4604

2018
[27]

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research21, 140 (2020), 1–67

2020
[28]

Jaejung Seol, Seojun Kim, and Jaejun Yoo. 2024. Posterllama: Bridging design ability of langauge model to contents-aware layout generation.arXiv preprint arXiv:2404.00995(2024)

2024 arXiv
[29]

Chameleon Team. 2024. Chameleon: Mixed-modal early-fusion foundation mod- els.arXiv preprint arXiv:2405.09818(2024)

2024 arXiv
[30]

Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang
[31]

Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie
[32]

Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. 2023. Layoutnuwa: Re- vealing the hidden layout expertise of large language models.arXiv preprint arXiv:2309.09506(2023)

2023 arXiv
[33]

Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. 2024. Diffusion model alignment using direct preference optimization. InProceedings of the IEEE/CVF Conference on Computer Visio...

2024
[34]

Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al
[35]

Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805(2023)

2023 arXiv
[36]

Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. 2023. Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis.arXiv preprint arXiv:2306.09341(2023)

2023 arXiv
[37]

Anytext: Multilingual visual text generation and editing.arXiv preprint arXiv:2311.03054(2023)

2023 arXiv
[38]

Ideogram v2

Ideogram AI. Ideogram v2. https://ideogram.ai/launch. 2024

2024
[39]

Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui
[40]

Zhendong Wang, Jianmin Bao, Shuyang Gu, Dong Chen, Wengang Zhou, and Houqiang Li. 2025. DesignDiffusion: High-Quality Text-to-Design Image Genera- tion with Diffusion Models.arXiv preprint arXiv:2503.01645(2025)

2025 arXiv
[41]

Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. 2024. Janus: Decoupling visual encoding for unified multimodal understanding and generation.arXiv preprint arXiv:2410.13848(2024)

2024 arXiv
[42]

bottom right

Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. 2024. Trans- fusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039(2024). 12•S...

2024 arXiv
[43]

Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. 2024. Show-o: One single transformer to unify multimodal understanding and generation.arXiv preprint arXiv:2408.12528(2024)

2024 arXiv
[44]

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

2025 arXiv
[46]

InForty-first International Conference on Machine Learning

Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. InForty-first International Conference on Machine Learning
[47]

Tao Yang, Yingmin Luo, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. 2024. PosterLLaVa: Constructing a Unified Multi-modal Layout Generator with LLM.arXiv preprint arXiv:2406.02884(2024)

2024 arXiv
[48]

Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. 2023. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22490–22499

2023
[2023]

arXiv preprint arXiv:2311.16465(2023)

Textdiffuser-2: Unleashing the power of language models for text rendering. arXiv preprint arXiv:2311.16465(2023)

2023 arXiv
[2024]

arXiv preprint arXiv:2411.15098(2024)

Ominicontrol: Minimal and universal control for diffusion transformer. arXiv preprint arXiv:2411.15098(2024)

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Stability AI. 2024. Stable Diffusion 3.5 Large. https://huggingface.co/stabilityai/ stable-diffusion-3.5-large

work page 2024

[2] [2]

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.arXiv preprint arXiv:2503.14908(2025)

arXiv 2025

[3] [4]

Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei

[4] [5]

Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan. 2025. Emerging Properties in Unified Multimodal Pretraining. arXiv:2505.14683 [cs.CV] https://arxiv.org/abs/2505.14683

arXiv 2025

[5] [6]

Textdiffuser: Diffusion models as text painters.Advances in Neural Infor- mation Processing Systems36 (2023), 9353–9387

2023

[6] [7]

Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. 2023. Layoutgpt: Compo- sitional visual planning and generation with large language models.Advances in Neural Information Processing Systems36 (2023), 18225–18250

work page 2023

[7] [8]

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. 2024. PosterCraft: Rethinking High-Quality Aesthetic Poster Generation in a Unified Framework•11 Scaling rectified flow transformers for high-resolution image synthesis. InForty- first International Confe...

work page 2024

[8] [9]

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

work page 2024