Dataset · kohsei
MultiBanana benchmark for multi-reference generation
kohsei/MultiBanana-Benchmark
A challenging CVPR 2026 test set for evaluating text-to-image generation from multiple references.
What's inside
This is the official dataset for the MultiBanana scientific benchmark accepted to the main track of CVPR 2026. It contains English images and text descriptions with 1 to 10 thousand records. The set is designed to test how well text-to-image models generate pictures based on multiple reference images simultaneously.
What you can do with it
- Evaluating and comparing image generation models using multiple references
- Testing character and object consistency in multimodal generators
- Tasks
- Text → image
- Languages
- English
- Size
- 1K–10K rows
- Downloads this month
- 26.9K
- Likes
- 5
How to load it
from datasets import load_dataset
ds = load_dataset("kohsei/MultiBanana-Benchmark")