PosterVerse: A Full-Workflow Framework for Commercial-Grade Poster Generation with HTML-Based Scalable Typography

AI summary

Overview

Research area: Computer vision and generative AI, specifically automated graphic design and text-to-image synthesis for poster generation.

Technical level: Advanced. The work combines fine-tuned large language models, LoRA-adapted diffusion models, and multimodal LLMs, and assumes familiarity with diffusion models, layout planning, and HTML rendering pipelines.

Scope: The paper presents PosterVerse, a three-stage, prompt-driven framework that generates commercial-grade, text-dense posters as editable HTML documents, along with PosterDNA, a 167,000-instance Chinese poster dataset with HTML-based typography annotations.

What This Paper Is About

Automated poster generation tools today fall short of commercial standards: they produce illegible or incorrect text (especially for Chinese characters and small/dense fonts), require awkward extra inputs like masks or bounding boxes, and output static images that cannot be edited afterward. This paper proposes PosterVerse, an end-to-end workflow that mimics how professional designers work and outputs HTML files instead of static images, achieving near-perfect text fidelity with scalable typography while remaining fully prompt-driven.

Key Contributions

1. PosterDNA dataset. The first commercial-grade, text-dense Chinese poster generation dataset with fine-grained HTML typography files, totaling 167,000 instances across three specialized subsets (57,000 blueprint-creation, 100,000 graphic-generation, 9,000 unified layout-text, plus a 1,000-sample test set). It is the first dataset to unify layout and text rendering in a single editable format.
2. PosterVerse framework. A full-workflow method integrating blueprint creation (fine-tuned LLM), graphical background generation (LoRA-tuned Flux.1-dev across four style variants), and unified layout-text rendering (fine-tuned Qwen2.5-VL-7B producing HTML).
3. HTML-based scalable typography. By rendering text through HTML rather than as pixels, the framework achieves 100% text fidelity and natively editable output, fundamentally solving the small, high-density text problem that plagues diffusion-based generators.
4. Prompt-only usability. Users need only a natural-language prompt—no masks, bounding boxes, or pre-supplied subjects—with the Detail-Insensitive Requirement Parsing (DIPR) mechanism ensuring consistent quality regardless of how vague or detailed the input is.

Main Findings

• Text accuracy dominates existing models. PosterVerse achieves a Correct Rate of 92.33% and F1 of 78.58%, surpassing the next best method (GPT-4o at 49.73% CR / 48.49% F1) by at least 42.42 points in CR and 30.09 points in F1.
• Superior perceptual and layout quality. The framework records an FID of 62.54 (best among all baselines) and an overlap score of 0.0027, indicating cleaner, less cluttered layouts.
• Strong human preference. In a user study with 30 participants, PosterVerse received 71% of votes for best overall impression and practicality, versus 24% for GPT-4o and 2% or less for all other methods.
• Robustness to prompt detail. The DIPR mechanism keeps performance stable across basic, medium, and detailed user requirements, whereas baselines like Seedream 3.0 and GPT-4o degrade sharply with brief prompts.
• Hierarchical prompt sampling improves background generation. Training with three prompt-detail levels (rather than only detailed prompts) cut FID from 136.72 to 62.54 and raised CLIP-IS from 62.39 to 77.85.
• Multilingual generalization. The model produces English, French, Japanese, German, and Arabic posters without language-specific training, simply by translating the requirement while preserving the target text language.
• Recognizable failure modes in competitors. Qualitative analysis identifies three recurring baseline errors: text rendering corruption, wrong information (e.g., mis-transcribed phone numbers), and missing information (e.g., omitted addresses).

Methodology in Plain English

The researchers broke poster design into three stages that mirror a professional designer's workflow.

First, a blueprint stage uses a fine-tuned Qwen2.5-14B LLM to convert any user prompt—vague or detailed—into a structured JSON specification containing text content (titles, contact info), background style attributes, and parameters like color, theme, and resolution. Training with deliberately varied prompt detail levels makes the model insensitive to input vagueness.

Second, a background stage generates the visual graphic by fine-tuning Flux.1-dev with LoRA across four aesthetic categories: Illustrative, Design-Oriented, Minimalistic, and Photorealistic. To improve output quality, training images are grouped by resolution and aspect ratio (bucketing), and prompts are sampled from three hierarchical detail levels. Users can also upload their own background.

Third, a layout-text stage feeds the blueprint and background into a fine-tuned Qwen2.5-VL-7B multimodal model that outputs a complete HTML document. Because HTML handles text as code rather than pixels, typography is always accurate and fully editable; rendering in a browser yields the final poster.

To train this pipeline, the team built PosterDNA by using Claude 3.7 Sonnet to reverse-engineer requirements and generate annotations from real posters, followed by manual correction by professional designers—a phase that consumed about 80% of the four-month annotation effort. Evaluation combined objective metrics (OCR-based accuracy, FID, overlap) with GPT-4o ratings and a human study.

Why This Matters

Impact on research: The paper reframes poster generation from "image synthesis" to "structured document generation," showing that HTML can serve as a bridge between generative models and the precision demands of professional design. The PosterDNA dataset provides a benchmark for a task that previously lacked standardized evaluation resources.

Real-world applications:

• Marketing and advertising teams generating campaign posters from brief text descriptions.
• Small businesses and event organizers producing professional posters without hiring designers.
• E-commerce and retail creating product promotion materials at scale with post-generation editing.
• Multilingual publishing and localization workflows where text accuracy and script support are critical.

Industry relevance: Because output is editable HTML, the framework fits naturally into existing web and design toolchains, enabling brand teams to adjust fonts, copy, and layout after generation rather than regenerating from scratch. This addresses the dynamic, revision-heavy reality of commercial design work that static-image generators cannot serve.

Future Directions

• Reducing generation time. The 2–3 minute per-poster latency is a practical bottleneck for high-volume commercial use.
• Tightening stage coupling. Occasional misalignment between background graphics and text layout in complex designs suggests the three stages could share feedback or be jointly optimized.
• Broadening style and domain coverage. The graphic-generation stage was trained on four fixed background styles; expanding to more diverse aesthetics and niche domains remains open.
• Richer evaluation of editability. The paper demonstrates HTML editability but does not formally measure how well post-edits preserve design quality, leaving room for dedicated metrics and user studies.

Target Audience

Researchers in computer vision, generative AI, and document analysis working on layout generation, visual text synthesis, or multimodal design automation. The paper is also valuable to industry practitioners in advertising, marketing technology, and e-commerce who need automated, editable, text-accurate design tools, and to dataset builders interested in HTML-based structured annotation for creative content.

Authors’ abstract

Commercial-grade poster design demands the seamless integration of aesthetic appeal with precise, informative content delivery. Current automated poster generation systems face significant limitations, including incomplete design workflows, poor text rendering accuracy, and insufficient flexibility for commercial applications. To address these challenges, we propose PosterVerse, a full-workflow, commercial-grade poster generation method that seamlessly automates the entire design process while delivering high-density and scalable text rendering. PosterVerse replicates professional design through three key stages: (1) blueprint creation using fine-tuned LLMs to extract key design elements from user requirements, (2) graphical background generation via customized diffusion models to create visually appealing imagery, and (3) unified layout-text rendering with an MLLM-powered HTML engine to guarantee high text accuracy and flexible customization. In addition, we introduce PosterDNA, a commercial-grade, HTML-based dataset tailored for training and validating poster design models. To the best of our knowledge, PosterDNA is the first Chinese poster generation dataset to introduce HTML typography files, enabling scalable text rendering and fundamentally solving the challenges of rendering small and high-density text. Experimental results demonstrate that PosterVerse consistently produces commercial-grade posters with appealing visuals, accurate text alignment, and customizable layouts, making it a promising solution for automating commercial poster design. The code and model are available at https://github.com/wuhaer/PosterVerse.