https://github.com/sony/banneragency
The official repo for BannerAgency: Advertising Banner Design with Multimodal LLM Agents
https://github.com/sony/banneragency
Last synced: 4 months ago
JSON representation
The official repo for BannerAgency: Advertising Banner Design with Multimodal LLM Agents
- Host: GitHub
- URL: https://github.com/sony/banneragency
- Owner: sony
- Created: 2025-07-01T04:58:47.000Z (5 months ago)
- Default Branch: main
- Last Pushed: 2025-08-04T04:09:06.000Z (4 months ago)
- Last Synced: 2025-08-04T07:27:52.573Z (4 months ago)
- Language: Python
- Homepage: https://banneragency.github.io/
- Size: 4.85 MB
- Stars: 2
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# BannerAgency: Advertising Banner Design with Multimodal LLM Agents
An autonomous agent system for editable banner ad image generation empowered by multiple multimodal LLMs.
**BannerAgency: Advertising Banner Design with Multimodal LLM Agents**
Authors: [Heng Wang](https://scholar.google.com.au/citations?user=jPj4ViQAAAAJ&hl=en), [Yotaro Shimose](https://github.com/yotaro-shimose), and [Shingo Takamatsu](https://scholar.google.co.jp/citations?user=oCVG8wQAAAAJ&hl=en) from Sony Group Corporation
## Installation
```bash
# Clone the repository
conda create -n baenv python=3.10
pip install -r requirements.txt
```
Please specify your API keys by copying `.env.example` to `.env`.
## Scheduled Releases
August 4, 2025
- [x] All the agent prompts have been released.
July 1, 2025
- [x] BannerRequest400 dataset release.
- [x] Evaluation code release.
## BannerRequest400 Benchmark
### Logo images
The 100 logos were generated semi-automatically with the assistance of [Claude-3.5 Sonnet](https://www.anthropic.com/news/claude-3-5-sonnet). We manualy checked the content to ensure quality. We release both the rendered `.png` format (`BannerRequest400/logos_png`) and the raw `.svg` format (`BannerRequest400/logos_svg`) to suit custom needs.
### Banner requests
Accompanying the 100 logo images are 400 abstract banner requests (`BannerRequest400/abstract_400.jsonl`) with four different target-purpose pairs for each logo. The abstract banner requests are derived from the [DESIGNERINTENTION](https://docs.google.com/spreadsheets/d/1e1-xFgPg_gzwy61xvsmtCpI_YjT8ofrYr9SqfthxgYg/edit?gid=0#gid=0) proposed in COLE [1]. We use [GPT-4o](https://openai.com/index/hello-gpt-4o/) for the target-purpose pair creation.
We also extend the 400 abstract banner requests to 5200 concrete requests across 13 standard banner dimensions with detailed banner specifications (`BannerRequest400/concrete_5k.jsonl`) via GPT-4o.
[1] Jia, Peidong, et al. "COLE: A Hierarchical Generation Framework for Multi-Layered and Editable Graphic Design." arXiv preprint arXiv:2311.16974 (2023).
## Evaluation
```bash
python3 eval.py --logo_file --image_file --banner_request
```
## Citation
If you find our work helpful in your research, please kindly cite our paper via:
```bibtex
@article{wang2025banneragency,
title = {BannerAgency: Advertising Banner Design with Multimodal LLM Agents},
author = {Wang, Heng and Shimose, Yotaro and Takamatsu, Shingo},
url = {arxiv},
year = {2025},
}
```
## Contact
If you have any questions or suggestions about this repo, please feel free to contact me ([heng.wang@sony.com](heng.wang@sony.com)) or Shingo Takamatsu ([shingo.takamatsu@sony.com](shingo.takamatsu@sony.com))