Zhao Zhang, Yutao Cheng, Dexiang Hong, Maoke Yang,
Gonglei Shi, Lei Ma, Hui Zhang, Jie Shao, and Xinglong Wu
[arXiv 📚]
[Model ⚙️]
[Results 🖼️]
[Bibtex 🔗]
This repository open-sources CreatiPoster, an AI-driven graphic design generation system that supports multi-layer and editable compositions with strong visual appeal.
Example of editing a graphic composition in the CreatiPoster editor. Users can modify text, assets, layout, and style through an intuitive GUI with JSON field controls, enabling professional-level customization.
[2024/06/13] Our manuscript is now available on arXiv.
- Publish the training and inference codebase
- Release the CreatiPoster-F model checkpoint
- Release the training dataset and evaluation set
Overview of the CreatiPoster pipeline: User inputs are processed by the protocol model to generate editable design layers, while the background model creates a complementary background. The final graphic composition integrates both outputs seamlessly.
CreatiPoster supports diverse interaction modes, including prompt-only, asset-only, mixed input, and explicit specification of text/asset layout or attributes.
CreatiPoster can extend static graphic compositions to animated posters. Videos are generated from background layers using an image-to-video model.
 |
 |
 |
|---|---|---|
 |
 |
 |
CreatiPoster's multilingual capabilities: Despite training data only including Simplified Chinese and English graphic compositions, pre-training and multilingual fine-tuning enable the protocol model to generalize to other languages.
Our protocol model enables direct text overlay on uploaded assets without needing the background model. This is ideal for tasks like adding titles to e-commerce product images or text overlays on social media photos.
Given an original graphic design, CreatiPoster generates alternative layouts of various sizes while preserving content and style. By reusing rendered layers and predicting new foreground/background elements, this approach enables efficient adaptation of designs for different platforms.
This system has been integrated into Pippit AI to power its poster generation capabilities.
