PosterCraft: A High-Quality Poster Generation Framework 253

小编 · 2025-06-24

PosterCraft is a poster generation framework jointly proposed in June 2025 by research teams from HKUST (Guangzhou), Meituan, Xiamen University, and the National University of Singapore. It combines precise text rendering and aesthetic style optimization with region-aware calibration and reinforcement learning to enhance the artistic completeness and overall harmony of generated posters. A vision-language feedback mechanism further refines aesthetic quality through multimodal corrections.

PosterCraft: A High-Quality Poster Generation Framework

PosterCraft is a poster generation framework jointly proposed in June 2025 by research teams from The Hong Kong University of Science and Technology (Guangzhou), Meituan, Xiamen University, and the National University of Singapore. Through precise text rendering and aesthetic style optimization, combined with region-aware calibration and reinforcement learning techniques, PosterCraft enhances the artistic integrity and overall harmony of posters. At the same time, it leverages a vision-language feedback mechanism to refine and optimize aesthetic effects through multimodal corrections.

In terms of dataset construction, PosterCraft has developed multiple high-quality datasets to meet diverse needs in text rendering and poster generation, laying the foundation for creating posters that align with human aesthetic standards.

PosterCraft：高质量海报生成框架 .io.webp

Its core features include:

Precise text rendering: Accurately presents diverse text on high-quality backgrounds, ensuring harmony and unity between text and background, and solving the problem of insufficient text rendering fidelity.
Aesthetic style optimization: Leverages region-aware calibration techniques to strengthen the artistic integrity of posters, maintain text accuracy, and build a harmonious visual relationship between text and background.
Aesthetic preference optimization: Through reinforcement learning techniques, optimizes outputs based on overall aesthetic standards, reduces font rendering defects, and enables the model to learn higher-level aesthetic preferences such as compositional balance and color harmony.
Vision-language feedback: With the help of a joint vision-language conditioning mechanism, combines visual information with targeted textual suggestions for multimodal corrections, iteratively optimizing poster aesthetic content and background coherence.

The workflow of PosterCraft is divided into four stages:

Text rendering optimization: Conducts large-scale training on the self-built Text-Render-2M dataset (containing 2 million high-quality text rendering examples, supporting multi-instance text rendering and diverse text selections) to improve the model's text rendering fidelity.
Poster style fine-tuning: Uses the HQ-Poster-100K dataset (a collection of 100,000 high-quality posters, employing advanced filtering techniques and multimodal scoring systems) together with a region-aware calibration strategy for fine-tuning, enabling the model to learn the harmonious coexistence of text and background.
Aesthetic text reinforcement learning: Adopts a best-of-n based DPO strategy, allowing the model to learn higher-level aesthetic preferences such as compositional balance and color harmony by comparing multiple versions of posters.
Vision-language feedback: Introduces VLM-based multimodal feedback to iteratively refine and correct content and style during the generation process, and allows post-generation adjustment and optimization based on design suggestions.

The technical advantages of this framework are reflected in:

Unified framework design: Abandoning the modular "planning-then-generation" pipeline, it achieves end-to-end generation from text input to poster output, ensuring visual coherence.
Automated dataset construction: Designs dedicated fully automated datasets for each optimization stage, addressing data scarcity issues.
Cascaded optimization strategy:

Large-scale text rendering optimization: Leverages high-quality datasets to focus on text rendering while preserving background rendering capability.
Region-aware calibration: Assigns different weights to text and non-text regions during fine-tuning, preserving text clarity while reinforcing artistic style.
Preference-based reinforcement learning: Learns from "best/worst" sample pairs to inject human aesthetic preferences.
Multimodal feedback loop: Incorporates multimodal generation correction suggestions during inference, integrating them into conditional inputs to iteratively optimize results.

Tests show that PosterCraft consistently surpasses existing state-of-the-art methods in text rendering accuracy, aesthetic quality, and layout structure, and remains competitive compared to commercial systems. The model can handle complex prompts while maintaining visual consistency, generating high-quality posters for both long texts and short prompts. Its advantage lies in directly generating posters from a single prompt, without requiring additional modules or manual adjustments, providing a fast and efficient scalable solution for movie posters, event promotions, product marketing, and other scenarios, without compromising aesthetic quality.