Repository

BannerAgency

BannerAgency: Advertising Banner Design with Multimodal LLM Agents An autonomous agent system for editable banner ad image generation empowered by multiple multimodal LLMs. <a href · License unknown · Install via pip · Maintenance unknown

License
Unknown

README

BannerAgency: Advertising Banner Design with Multimodal LLM Agents

An autonomous agent system for editable banner ad image generation empowered by multiple multimodal LLMs.

<br>
<a href="https://arxiv.org/pdf/2503.11060">

</a>
<a href="https://banneragency.github.io/">

</a>

BannerAgency: Advertising Banner Design with Multimodal LLM Agents, EMNLP 2025 Main

Authors: Heng Wang, Yotaro Shimose, and Shingo Takamatsu from Sony Group Corporation

Installation

# 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. 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.

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 proposed in COLE [1]. We use 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

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:

@article{wang2025banneragency,
  title     = {BannerAgency: Advertising Banner Design with Multimodal LLM Agents},
  author    = {Wang, Heng and Shimose, Yotaro and Takamatsu, Shingo},
  url = {EMNLP Main},
  year      = {2025},
}

Contact

If you have any questions or suggestions about this repo, please feel free to contact me ([email protected]) or Shingo Takamatsu ([email protected])