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

🌐 Website | 🎯 Demo | 📄 Paper | 🤗 Models | 📚 Datasets | 🎥 Video | 🤗 HF Demo

• 🧩 [2025.06] Community user @AIFSH has successfully integrated PosterCraft into ComfyUI!
  You can check out the full workflow here: PosterCraft-ComfyUI Example
  Big thanks to the contributor — this will be helpful for many users! See Issue #6 for details.
• 📖 [2025.06] Our Chinese article providing a detailed introduction and technical walkthrough of PosterCraft is now available!
  Read it here: 中文解读｜高质量美学海报生成框架 PosterCraft
• 🔥 [2025.06] We have deployed a demo on Hugging Face Space, feel free to give it a try!
• 🚀 [2025.06] Our gradio demo and inference code are now available!
• 📊 [2025.06] We have released partial datasets and model weights on HuggingFace.

Sixiang Chen^1,2*, Jianyu Lai¹*, Jialin Gao²*, Tian Ye¹, Haoyu Chen¹, Hengyu Shi², Shitong Shao¹, Yunlong Lin³, Song Fei¹, Zhaohu Xing¹, Yeying Jin⁴, Junfeng Luo², Xiaoming Wei², Lei Zhu^1,5†

¹The Hong Kong University of Science and Technology (Guangzhou)
²Meituan
³Xiamen University
⁴National University of Singapore
⁵The Hong Kong University of Science and Technology

*Equal Contribution, †Corresponding Author

PosterCraft is a unified framework for high-quality aesthetic poster generation that excels in precise text rendering, seamless integration of abstract art, striking layouts, and stylistic harmony.

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

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

# Install dependencies
pip install -r requirements.txt

Generate high-quality aesthetic posters from your prompt with BF16 precision:

python inference.py \
  --prompt "Urban Canvas Street Art Expo poster with bold graffiti-style lettering and dynamic colorful splashes" \
  --enable_recap \
  --num_inference_steps 28 \
  --guidance_scale 3.5 \
  --seed 42 \
  --pipeline_path "black-forest-labs/FLUX.1-dev" \
  --custom_transformer_path "PosterCraft/PosterCraft-v1_RL" \
  --qwen_model_path "Qwen/Qwen3-8B"

If you are running on a GPU with limited memory, you can use inference_offload.py to offload some components to the CPU:

python inference_offload.py \
  --prompt "Urban Canvas Street Art Expo poster with bold graffiti-style lettering and dynamic colorful splashes" \
  --enable_recap \
  --num_inference_steps 28 \
  --guidance_scale 3.5 \
  --seed 42 \
  --pipeline_path "black-forest-labs/FLUX.1-dev" \
  --custom_transformer_path "PosterCraft/PosterCraft-v1_RL" \
  --qwen_model_path "Qwen/Qwen3-8B"

We provide a Gradio web UI for PosterCraft.

python demo_gradio.py

📈 Quantitative Results

Method	Text Recall ↑	Text F-score ↑	Text Accuracy ↑
OpenCOLE (Open)	0.082	0.076	0.061
Playground-v2.5 (Open)	0.157	0.146	0.132
SD3.5 (Open)	0.565	0.542	0.497
Flux1.dev (Open)	0.723	0.707	0.667
Ideogram-v2 (Close)	0.711	0.685	0.680
BAGEL (Open)	0.543	0.536	0.463
Gemini2.0-Flash-Gen (Close)	0.798	0.786	0.746
PosterCraft (ours)	0.787	0.774	0.735

🎨 PosterCraft Gallery

Adventure Travel	Post-Apocalyptic	Sci-Fi Drama
Space Thriller	Cultural Event	Luxury Product
Concert Show	Children's Book	Movie Poster

A unified framework for high-quality aesthetic poster generation

Our unified framework consists of four critical optimization stages in the training workflow:

Addresses accurate text generation by precisely rendering diverse text on high-quality backgrounds, also ensuring faithful background representation and establishing foundational fidelity and robustness for poster generation.

Shifts focus to overall poster style and text-background harmony using Region-aware Calibration. This fine-tuning stage preserves text accuracy while strengthening the artistic integrity of the aesthetic poster.

Employs Aesthetic-Text Preference Optimization to capture higher-order aesthetic trade-offs. This reinforcement learning stage prioritizes outputs that satisfy holistic aesthetic criteria and mitigates defects in font rendering.

Introduces a Joint Vision-Language Conditioning mechanism. This iterative feedback combines visual information with targeted text suggestions for multi-modal corrections, progressively refining aesthetic content and background harmony.

We provide the weights for our core models, fine-tuned at different stages of the PosterCraft pipeline.

We provide four specialized datasets for training PosterCraft workflow:

Text-Render-2M: Multi-instance text rendering with diverse selections

A comprehensive text rendering dataset containing 2 million high-quality examples. Features multi-instance text rendering, diverse text selections (varying in size, count, placement, and rotation), and dynamic content generation through both template-based and random string approaches.

HQ-Poster-100K: Curated high-quality aesthetic posters

100,000 meticulously curated high-quality posters with advanced filtering techniques and multi-modal scoring. Features Gemini-powered mask generation with detailed captions for comprehensive poster understanding.

Poster-Preference-100K: Preference learning pairs for aesthetic optimization

This preference dataset is sourced from over 100,000 generated poster images. Through comprehensive evaluation by Gemini and aesthetic evaluators, we construct high-quality preference pairs designed for reinforcement learning to align poster generation with human aesthetic judgments.

Poster-Reflect-120K: Vision-language feedback pairs for iterative refinement

This vision-language feedback dataset is sourced from over 120,000 generated poster images. Through comprehensive evaluation by Gemini and aesthetic evaluators, this dataset captures the iterative refinement process and provides detailed feedback for further improvements.

If you find PosterCraft useful for your research, please cite our paper:

@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}
}

• 🏛️ Thanks to our affiliated institutions for their support.
• 🤝 Special thanks to the open-source community for inspiration.

For any questions or inquiries, please reach out to us:

• Sixiang Chen: [email protected]
• Jianyu Lai: [email protected]