This repo provides a minimal, runnable artifact accompanying the paper.

This repository contains the official implementation of the multi-agent, multi-modality LLM system described for ad banner generation in the paper “Mirror in the Model: Ad Banner Image Generation via Reflective Multi-LLM and Multi-modal Agents” by Zhao Wang, Bowen Chen, Yotaro Shimose, Sota Moriyama, Heng Wang, Shingo Takamatsu. For simplicity, we refer to our model as MIMO. MIMO combines a hierarchical multi-modal agent system (MIMO-Core) with a coordination loop (MIMO-Loop) that explores multiple stylistic directions and iteratively improves design quality. Requiring only a simple natural language based prompt and logo image as input, MIMO automatically detects and corrects multiple types of errors during generation.

• Source code: src/multiagent, src/prompts, src/okg, src/result_manager.py
• Configuration: config/config_llm.ini (temporarily includes keys for reproduction; remove or switch to env vars after verification)
• Example data: logos/001_ethicai.png and logos/001_ethicai_prompt.txt
• Entry script: test_naming_convention.py
• One-click run script: scripts/run_example.sh

1. Run the example (creates a venv and installs deps on first run)

bash scripts/run_example.sh

2. Run with a different logo

python test_naming_convention.py --logo wildcare

See available logos in logos/.

3. Dynamic styles (optional)

python test_naming_convention.py --logo ethicai --dynamic-styles --style-candidates 8

Final images follow:

generated_image_GraphicRevisor_Team{TeamID}_Round{RoundNumber}_{RevisionNumber}.png

The script prints verification details in the console.

• Main configuration: config/config_llm.ini
• Image model is configured via IMAGE_GENERATION.MODEL (e.g., gemini or gpt)
• Keys are temporarily present under [KEYS] for reproduction; after verifying the run, please remove them or use environment variables instead

• GEMINI_API_KEY
• If switching to OpenAI/Azure paths, the corresponding key and endpoint variables

• Python 3.10+
• See requirements.txt

paper_artifact/
  ├── README.md
  ├── requirements.txt
  ├── scripts/
  │   └── run_example.sh
  ├── config/
  │   └── config_llm.ini
  ├── logos/
  │   ├── 001_ethicai.png
  │   └── 001_ethicai_prompt.txt
  ├── src/
  │   ├── multiagent/
  │   ├── prompts/
  │   ├── okg/
  │   └── result_manager.py
  └── test_naming_convention.py

• Keys in the example configuration are only for paper reproduction. Do not commit them to public repositories. After confirming local runs, remove keys or switch to environment variables.

This project is licensed under the MIT License.

Contributions are welcome! Please fork the repository and submit a pull request for any improvements or bug fixes.

For any questions or issues, feel free to reach out: Zhao.Wang@sony.com or this github repo for any information.

If you use or reference TalkHier, please cite us with the following BibTeX entry:

@inproceedings{Wang_etal_2025_MIMO,
  title     = {Mirror in the Model: Ad Banner Image Generation via Reflective Multi-LLM and Multi-modal Agents},
  author    = {Wang, Zhao and Chen, Bowen and Shimose, Yotaro and Moriyama, Sota and Wang, Heng and Takamatsu, Shingo},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2025},
  publisher = {Association for Computational Linguistics},
  url       = {https://arxiv.org/abs/2507.03326}
}