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

The Design Revolution You’ve Been Waiting For

Have you ever struggled to create professional posters? Faced with fuzzy text rendering in AI-generated designs? Watched artistic elements clash with backgrounds? PosterCraft solves these challenges through its groundbreaking unified framework. Developed collaboratively by researchers from The Hong Kong University of Science and Technology, Meituan, Xiamen University, and National University of Singapore, this innovative system achieves unprecedented precision in text rendering and aesthetic harmony.

Performance breakthrough: PosterCraft achieves 0.787 text recall – outperforming SD3.5 (0.565) and nearly matching Gemini2.0 (0.798) in independent benchmarks.

What Makes PosterCraft Special?

PosterCraft is the first unified framework dedicated to high-quality aesthetic poster generation. It combines four core capabilities in a single architecture:

1. Precision text rendering – Crystal-clear readable text elements
2. Seamless artistic integration – Natural blending of abstract elements
3. Dynamic layout composition – Professional-grade visual arrangements
4. Style consistency – Unified aesthetic across all components

Performance Comparison

Model	Text Recall	Text Accuracy	Aesthetic Score
SD3.5	0.565	0.497	3.2/5
Flux1.dev	0.723	0.667	3.8/5
PosterCraft	0.787	0.735	4.5/5

The Four-Stage Technical Architecture

Stage 1: Text Rendering Optimization

[object Promise]

• ☾ Trained on 2 million high-quality samples
• ☾ Supports variable text size, position, and rotation
• ☾ Solves core issue of fuzzy text in generative models

Stage 2: Poster Fine-Tuning

[object Promise]

• ☾ Uses 100,000 curated poster examples
• ☾ Maintains text precision while boosting artistic quality
• ☾ Eliminates visual dissonance between elements

Stage 3: Aesthetic-Text Reinforcement

[object Promise]

• ☾ Trained on 100,000 preference datasets
• ☾ Incorporates Gemini model aesthetic scoring
• ☾ Achieves high-level aesthetic-text balance

Stage 4: Vision-Language Feedback

[object Promise]

• ☾ Utilizes 120,000 feedback samples
• ☾ Combines visual and textual correction
• ☾ Progressively improves detail quality

Getting Started in 5 Minutes

Installation Guide

# Clone repository
git clone https://github.com/ephemeral182/PosterCraft.git
cd PosterCraft

# Create environment
conda create -n postercraft python=3.11
conda activate postercraft

# Install dependencies
pip install -r requirements.txt

Generate Your First Poster

python inference.py \
  --prompt "Urban street art expo poster with graffiti text and dynamic color splashes" \
  --enable_recap \
  --num_inference_steps 28 \
  --guidance_scale 3.5 \
  --seed 42

Solution for Limited GPU Memory

python inference_offload.py \
  --prompt "Sci-fi movie poster with spaceship and nebula background" \
  --enable_recap

Visual Interface

python demo_gradio.py

Access the interactive interface through your local port after launch

Gallery of Generated Posters

Category	Example	Prompt Keywords
Adventure Travel		Natural landscapes + dynamic typography
Post-Apocalyptic		Decayed textures + bold titles
Sci-Fi Drama		Neon colors + tech elements
Cultural Event		Traditional patterns + gold accents
Children’s Book		Soft colors + rounded fonts

Pro tip: Including terms like “bold lettering” or “dynamic layout” in prompts significantly enhances results

Models and Datasets

Core Models

Model	Training Stage	Key Features	Download
PosterCraft-v1_RL	Stage 3	Aesthetic-text balance	🤗 Hugging Face
PosterCraft-v1_Reflect	Stage 4	Multimodal optimization	🤗 Hugging Face

Training Datasets

Dataset	Size	Purpose	Features
Text-Render-2M	2M samples	Text rendering	Multi-instance support Download
HQ-Poster-100K	100K samples	Poster refinement	Professional curation Download
Poster-Preference-100K	100K samples	RL training	Aesthetic preference pairs Download
Poster-Reflect-120K	120K samples	Feedback training	Multimodal analysis Download

Practical Questions Answered

What are the GPU requirements?

• ☾ Standard mode: Recommended 12GB+ VRAM
• ☾ Limited VRAM solution: Use inference_offload.py with 8GB VRAM

Is Chinese language supported?

Yes! Recent developments include:

[June 2025] Chinese technical article released: Detailed Introduction to PosterCraft

What are the licensing terms?

Current version uses open-source licensing. For commercial licensing, contact the authors directly.

How to integrate with existing workflows?

Community integration with ComfyUI already available:

[object Promise]

Full workflow: PosterCraft-ComfyUI Example

Academic Recognition

Core Research Institutions:

• ☾ 🏛️ The Hong Kong University of Science and Technology (Guangzhou)
• ☾ 🏢 Meituan
• ☾ 🏫 Xiamen University
• ☾ 🌏 National University of Singapore

Citation:

@article{chen2025postercraft,
  title={PosterCraft: Rethinking High-Quality Aesthetic Poster Generation in a Unified Framework},
  author={Chen, Sixiang and Lai, Jianyu and Gao, Jialin and Ye, Tian and Chen, Haoyu and Shi, Hengyu and Shao, Shitong and Lin, Yunlong and Fei, Song and Xing, Zhaohu and Jin, Yeying and Luo, Junfeng and Wei, Xiaoming and Zhu, Lei},
  journal={arXiv preprint arXiv:2506.10741},
  year={2025}
}

Connect With the Team

• ☾ Sixiang Chen: shen691@connect.hkust-gz.edu.cn
• ☾ Jianyu Lai: jlai218@connect.hkust-gz.edu.cn

Stay updated:
Official Website |
Research Paper |
Video Demo |
Live Experience