PosterCraft: Unified Poster Synthesis

• PosterCraft is a unified, data-driven AI framework for fully generative poster synthesis that integrates text rendering, layout optimization, and aesthetic control in an end-to-end diffusion transformer model.
• It employs a four-stage cascaded training workflow with specialized datasets and multimodal objectives to refine style, text fidelity, and dynamic content integration.
• The framework outperforms traditional modular methods in quantitative metrics and user evaluations, supporting interactive editing, multilingual adaptation, and flexible design applications.

PosterCraft is a unified, data-driven AI framework designed for high-fidelity, flexible, and fully generative poster synthesis, with particular emphasis on aesthetic quality, robust text rendering, and layout coherence. Unlike modular or template-driven pipelines, PosterCraft abandons rigid, pre-defined layouts in favor of an integrated cascade of dataset-engineered training stages and multimodal optimization objectives. Its open, staged design philosophy enables both end-to-end poster generation and fine-grained applications such as interactive editing, multilingual adaptation, and dynamic content integration (Chen et al., 12 Jun 2025).

1. Paradigm Shift: Unified Generative Poster Design

PosterCraft rejects the prevailing “modular + rigid-layout” paradigm—where text placement, layout, and image generation are handled with separate vision-language modules and diffusion backgrounds—because such decomposition leads to aesthetic fragmentation, restricted creative space, and susceptibility to module-level error propagation. Instead, PosterCraft uses a fully unified, cascaded workflow: all aesthetic, textual, and compositional factors are successively optimized on a single diffusion transformer backbone (Flux1.dev), fusing layout control, text fidelity, and visual harmony in an end-to-end process (Chen et al., 12 Jun 2025).

The inference interface requires only a single prompt (optionally LLM-augmented); the system autonomously determines text rendering, color balance, composition, and style. This architecture, echoing strategies in DreamPoster (Hu et al., 6 Jul 2025) and insights from GlyphDraw2’s triple-cross attention (Ma et al., 2024), enables both superior quantitative outcomes and a higher ceiling for creativity.

2. Four-Stage Cascaded Workflow

PosterCraft’s core innovation is a cascaded four-stage training strategy, each supported by automated dataset pipelines:

1. Text-Rendering Optimization: Fine-tuning on “Text-Render-2M,” a dataset of 2 million synthetic images with varied crisp text placements, via flow matching loss:

$$\mathcal{L}^{\rm text}_{\rm flow}(\phi) = \mathbb{E}_{t\sim U(0,1),\,x_0,\,\varepsilon} \bigl\|v_{\phi}(x_t,t) - \dot{x}_t\bigr\|_2^2,$$

where $v_\phi$ predicts the velocity field for text-centric image samples.

2. Region-Aware Supervised Fine-Tuning: Leveraging HQ-Poster-100K (strictly filtered for visual/aesthetic quality), training with loss terms upweighted for major/minor text regions vs. non-text:

$$w(p)= \begin{cases} 0.6 & p\in\mathrm{LargeTextMask},\ 0.2 & p\in\mathrm{SmallTextMask},\ 1.0 & \text{otherwise}. \end{cases}$$

This ensures legibility and seamless text-background integration.

3. Aesthetic-Text Reinforcement Learning: Using Poster-Preference-100K, DPO-based objective applied for “best-of-n” preference optimization on both text fidelity and global aesthetics:

$$\mathcal{L}_{\rm RL}(\theta) = -\mathbb{E}_{c}\Bigl[ \log\sigma\bigl(\beta(\log\frac{p_\theta(x^+\!\mid c)}{p_{\rm ref}(x^+\!\mid c)} -\log\frac{p_\theta(x^-\!\mid c)}{p_{\rm ref}(x^-\!\mid c)})\bigr) \Bigr].$$

4. Joint Vision-Language Feedback Refinement: Using Poster-Reflect-120K, Gemini-annotated feedback (content and style) is encoded together with the prompt and image-level features, and optimized with a conditional flow-matching objective. This stage provides iterative, multimodal, bilingual critique, directly driving semantic and stylistic refinement.

Each training stage uses dedicated, pipeline-constructed data—synthetic for explicit disentanglement, web-scraped and filtered for style, best-of-n preference for RL, and multimodal feedback for late refinement.

3. Model Architecture, Representational Design, and Training

PosterCraft’s backbone is a diffusion transformer (Flux-dev), with LoRA adapters introduced for rapid and isolated fine-tuning during RL and joint vision-language feedback stages. No architectural modifications—such as explicit layout tokens or external overlays—are made to the core U-Net or transformer blocks, maximizing compatibility and transferability with new diffusion models.

Key hyperparameters and settings include:

• Mixed-precision training throughout.
• Adafactor/AdamW optimizers, staged learning-rate schedules.
• RL and VL-refinement stages with LoRA ranks 64 and 128, respectively.
• Stage-specific dataset construction, synthetic augmentation, multimodal annotation.

The design is inspired by broader trends in generative graphic modeling, notably the transformer–diffusion hybrids of DreamPoster (Hu et al., 6 Jul 2025), the editable multi-layer protocols of CreatiPoster (Zhang et al., 12 Jun 2025), and the data-efficient augmentation of Scan-and-Print (Hsu et al., 27 May 2025).

4. Quantitative and Qualitative Evaluation

PosterCraft demonstrates marked advances against both open-source and proprietary baselines. Relevant quantitative metrics are reported on 300 prompt sets (using Gemini2.5-Flash OCR):

Method	Recall ↑	F-score ↑	Accuracy ↑
OpenCOLE (Open)	0.082	0.076	0.061
SD3.5 (Open)	0.565	0.542	0.497
Flux1.dev (Open)	0.723	0.707	0.667
Ideogram-v2 (Closed)	0.711	0.685	0.680
BAGEL (Open)	0.543	0.536	0.463
Gemini2.0-Flash-Gen	0.798	0.786	0.746
PosterCraft (Ours)	0.787	0.774	0.735

PosterCraft consistently outperforms Flux1.dev (+6.4 recall, +6.7 F-score, +6.8 accuracy), and is competitive with leading closed commercial systems. In user studies with 20 professional designers and in LLM-based judgements, PosterCraft led on aesthetics, prompt alignment, and text fidelity. Qualitative analysis shows that PosterCraft maintains genre-specific style fidelity, robust long-prompt handling, and error-free text rendering.

5. Data Construction: Datasets and Automated Pipelines

PosterCraft’s success derives in part from meticulous dataset engineering. Its data pipeline is fully automated:

• Text-Render-2M: Synthetic, randomized text content, font, and layout arrangements.
• HQ-Poster-100K: Web-scraped, de-duplicated, deep-filtered by multi-modal LLMs and OCR; annotated with Gemini-generated captions, area-masked for text hierarchy.
• Poster-Preference-100K: Generated via best-of-n sampling, curated by HPSv2 and verified text correctness.
• Poster-Reflect-120K: Iterative best-of-6 outputs, each re-annotated for content and style feedback in multilingual, multimodal channels.

This content-centric approach to dataset curation ensures high training fidelity even in the absence of human-in-the-loop supervision, aligning with best practices from recent advances in semantic data augmentation and layout modeling (Hsu et al., 27 May 2025).

6. Applications, Generalization, and Comparative Context

PosterCraft’s unified model and editable outputs support a wide gamut of design scenarios:

• One-shot text+style poster generation for advertising, events, or artistry.
• Direct support for content overlays, multi-asset layouts, and multimodal prompts.
• Responsive resizing and canvas adaptation; multilingual typography and composition.
• Animation and integration with video generation (compositing static design atop video frames).

Comparative analyses highlight the following strengths:

• Surpasses general I2I and template-driven generation in layout flexibility, subject preservation, and typographic detail (Hu et al., 6 Jul 2025).
• Outperforms prior content-aware and weakly-supervised pipeline strategies (Text2Poster, AutoPoster) on both quantitative and user-facing design axes (Jin et al., 2023, Lin et al., 2023).
• Directly supports downstream interactive or agent-based editing paradigms through its clean, protocol-friendly output structure, as demanded by recent multi-agent frameworks (Sun et al., 21 May 2025, Zhang et al., 24 Aug 2025, Shi et al., 8 Jan 2026).

7. Limitations and Future Directions

PosterCraft, while architecturally extensible, is not without open challenges:

• Reliance on a pre-trained foundation model (such as flux1.dev) risks inheriting any pretraining biases or gaps.
• Artwork and style filtering are based on LLM-in-the-loop evaluation and annotation, introducing potential feedback artifacts.
• Coverage for extreme or avant-garde poster genres, especially those diverging from strong grid or typographic conventions, may remain suboptimal due to dataset bias.
• Full real-time interactivity—e.g., fine-grained element manipulation or design-in-the-loop—is a future extension point.

Future work will involve porting the pipeline to stronger diffusion backbones, expanding linguistic/generalization capacity (especially for complex scripts and dense layouts), and integrating human-in-the-loop or reinforcement based design refinements. Further scaling of editable protocol outputs, semantic feedback loops, and high-fidelity dataset construction remains an active area for research and system engineering (Chen et al., 12 Jun 2025, Hu et al., 6 Jul 2025, Sun et al., 21 May 2025, Shi et al., 8 Jan 2026).

---

References

• “PosterCraft: Rethinking High-Quality Aesthetic Poster Generation in a Unified Framework” (Chen et al., 12 Jun 2025)
• “DreamPoster: A Unified Framework for Image-Conditioned Generative Poster Design” (Hu et al., 6 Jul 2025)
• “GlyphDraw2: Automatic Generation of Complex Glyph Posters with Diffusion Models and LLMs” (Ma et al., 2024)
• “CreatiPoster: Towards Editable and Controllable Multi-Layer Graphic Design Generation” (Zhang et al., 12 Jun 2025)
• “P2P: Automated Paper-to-Poster Generation and Fine-Grained Benchmark” (Sun et al., 21 May 2025)
• “PosterGen: Aesthetic-Aware Paper-to-Poster Generation via Multi-Agent LLMs” (Zhang et al., 24 Aug 2025)
• “AutoPoster: A Highly Automatic and Content-aware Design System for Advertising Poster Generation” (Lin et al., 2023)
• “Text2Poster: Laying out Stylized Texts on Retrieved Images” (Jin et al., 2023)
• “Scan-and-Print: Patch-level Data Summarization and Augmentation for Content-aware Layout Generation in Poster Design” (Hsu et al., 27 May 2025)
• “Learning to Generate Posters of Scientific Papers by Probabilistic Graphical Models” (Qiang et al., 2017)
• “APEX: Academic Poster Editing Agentic Expert” (Shi et al., 8 Jan 2026)