🍌 CVPR 2026 (Main) 🍌

MultiBanana broadly covers problems specific to multi-reference settings, including varying the number of references (up to 8), domain and scale mismatches among references (e.g., photo vs. anime), and multilingual text rendering, and the presence of rare concepts (e.g., red banana).

🥇 Leaderboard

MultiBanana-Bench comprises 3,769 tasks across 36 task types designed to evaluate how well image generation models can faithfully incorporate information from multiple reference images. We report evaluation scores using Qwen3-VL-8B-Instruct, a fixed, open-weight judge model (GPT and Gemini are also supported as judges). We hope this benchmark, along with its evaluation framework using an open-source VLM as a judge, will serve as a foundation for future research in multi-reference text-to-image generation.

📦 Dataset

The data structure at the Hugging Face dataset is as follows.

data/
├── 3_back/
│   ├── 006_0.jpg
│   ├── 006_1.jpg
│   ├── 006_2.jpg
│   ├── 006_prompt.txt
│   ├── 014_0.jpg
│   ├── 014_1.jpg
│   ├── 014_2.jpg
│   ├── 014_prompt.txt
│   └── ...
├── 3_global/
│   └── ...
├── 3_local/
│   └── ...
└── ...

Download MultiBanana dataset by

git clone https://huggingface.co/datasets/kohsei/MultiBanana-Benchmark ./data

🛠️ Setup

git clone git@github.com:matsuolab/multibanana.git
cd multibanana
conda create -n multibanana python=3.12
conda activate multibanana
pip install -r requirements.txt

🧪 Evaluation

Generated images are expected to be saved in the same directory with the _generated suffix.

data/
├── 3_back/
│   ├── 006_0.jpg
│   ├── 006_1.jpg
│   ├── 006_2.jpg
│   ├── 006_prompt.txt
│   ├── 006_generated.jpg
│   ├── 014_0.jpg
│   ├── 014_1.jpg
│   ├── 014_2.jpg
│   ├── 014_prompt.txt
│   ├── 014_generated.jpg
│   └── ...
├── 3_global/
│   └── ...
├── 3_local/
│   └── ...
└── ...

We use gemini-2.5-flash via the Google GenAI SDK, and gpt-5-2025-08-07 via the OpenAI SDK.

Please set your API key in .env as follows

OPENAI_API_KEY=...
GEMINI_API_KEY=...

Run

# Gemini
python judge.py --base_dir ./data --model gemini --batch_size 32 --output_dir ./results
# GPT
python judge.py --base_dir ./data --model gpt --batch_size 32 --output_dir ./results

This will evaluate all generated images and save the results in {number}_{model}_judge.txt files (e.g., 006_gemini_judge.txt).

We also provide an evaluation script based on the open-source model Qwen3-VL as an alternative option. To run the script, you need to install transformers and flash-attn.

python qwenvl_judge.py --base_dir ./data --output_dir ./results

🏷️ Annotation

The dataset released on Hugging Face includes the following annotation files:

Task Difficulty Categories

Each task directory contains types.json. This file provides a dictionary mapping each set to its assigned difficulty category.

The category labels are defined as follows:

• domain: cross-domain
• scale: scale and viewpoint differences
• rare: rare concept
• ling: multilingual

Sets containing text that are not multilingual are labeled font.

Source of Reference Images

from_where.csv contains metadata indicating whether each reference image originates from a real dataset or was synthetically generated.

📄 License

Creative Commons Attribution Non Commercial 4.0

🙏 Acknowledgement

MultiBanana would not have been possible without the open research ecosystem around image generation.

We thank the LAION team for openly releasing LAION-5B, a subset of whose images this benchmark partially incorporates. Our evaluation framework relies on Qwen3-VL as a fixed, open-weight judge model, and we are grateful to the Qwen team for making capable vision-language models freely available.

We appreciate the teams behind the models we evaluate. Among proprietary systems, our study covers Nano Banana (Gemini 2.5 Flash Image) from Google DeepMind, and GPT-Image-1 from OpenAI. Among open-weight models, we evaluate Qwen-Image-Edit, FLUX.1 Kontext [dev], OmniGen2, and DreamOmni2.

Finally, MultiBanana builds on a substantial body of prior work in subject-driven and personalized image generation, including DreamBooth, IP-Adapter, and OmniGen. We also acknowledge earlier benchmarking efforts that shaped how this field is evaluated — MagicBrush and ImgEdit for instruction-based editing, and OmniContext for in-context, multi-reference generation.

🌟 Citation

@inproceedings{oshima2026multibanana,
    author    = {Oshima, Yuta and Miyake, Daiki and Matsutani, Kohsei and Iwasawa, Yusuke and Suzuki, Masahiro and Matsuo, Yutaka and Furuta, Hiroki},
    title     = {MultiBanana: A Challenging Benchmark for Multi-Reference Text-to-Image Generation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {448-460}
}