Pico-Banana-400K is a large-scale dataset of ~400K text–image–edit triplets designed to advance research in text-guided image editing.
Each example contains:

• an original image (from Open Images),
• a human-like edit instruction, and
• the edited result generated and verified by the Nano-Banana model.

The dataset spans 35 edit operations across 8 semantic categories, covering diverse transformations—from low-level color adjustments to high-level object, scene, and stylistic edits.

Pico-Banana-400K is built using a two-stage multimodal generation pipeline:

1. Instruction Generation
   Each Open Images sample is passed to Gemini-2.5-Flash, which writes concise, natural-language editing instructions grounded in visible content. We also provide short instructions summarized by Qwen-2.5-Instruct-7B. Example:

   {
     "instruction": "Change the red car to blue."
   }
   
   

2. Editing + Self-Evaluation The Nano-Banana model performs the edit, then automatically evaluates the result using a structured quality prompt that measures: Instruction Compliance (40%) Editing Realism (25%) Preservation Balance (20%) Technical Quality (15%) Only edits scoring above a strict threshold (~0.7) are labeled as successful, forming the main dataset; the remaining ~56K are retained as failure cases for robustness and preference learning.

Nano-Banana-400K contains ~400K image editing data, covering a wide visual and semantic range drawn from real-world imagery.

• Single-Turn SFT samples (successful edits): ~257K
• Single-Turn Preference samples (failure cases): ~56K
• Multi-Turn SFT samples (successful cases): ~72K
• Gemini-generated instructions: concise, natural, and image-aware
• Edit coverage: 35 edit types across 8 semantic categories
• Image diversity: includes humans, objects, text-rich scenes, etc from Open Images

Below are representative examples from different categories:

Pico-Banana-400K provides both breadth (diverse edit operations) and depth (quality-controlled multimodal supervision), making it a strong foundation for training and evaluating text-guided image editing models.

Pico-Banana-400K serves as a versatile resource for advancing controllable and instruction-aware image editing.
Beyond single-step editing, the dataset enables multi-turn, conversational editing and reward-based training paradigms.

The Pico-Banana-400K dataset is hosted on Apple’s public CDN.
You can download each component (single-turn, multi-turn, and preference data) using the provided manifest files.

Manifest files: sft link and preference link

Manifest file: multi-turn link

Urls to download source images are provided along with edit instructions in sft link, preference link, and multi-turn link. If you hit rate limit with Flickr when downloading images, you can either request higher rate limit with Flickr or follow steps below.

Another way to download the source images is to download packed files train_0.tar.gz and train_1.tar.gz from Open Images, then map with the urls we provide. We also provide a sample mapping code here. Due to legal requirements, we cannot provide the source image files directly.

# install awscli(https://aws.amazon.com/cli/)
# Download Open Images packed files 
aws s3 --no-sign-request --endpoint-url https://s3.amazonaws.com cp s3://open-images-dataset/tar/train_0.tar.gz . 
aws s3 --no-sign-request --endpoint-url https://s3.amazonaws.com cp s3://open-images-dataset/tar/train_1.tar.gz . 

# Create folder for extracted images 
mkdir openimage_source_images

# Extract the tar files 
tar -xvzf train_0.tar.gz -C openimage_source_images
tar -xvzf train_1.tar.gz -C openimage_source_images

# Download metadata CSV (ImageID ↔ OriginalURL mapping)  
wget https://storage.googleapis.com/openimages/2018_04/train/train-images-boxable-with-rotation.csv

# Map urls to local paths
python map_openimage_url_to_local.py #please modify variable is_multi_turn and file paths as needed

Pico-Banana-400K is released under the Creative Commons Attribution–NonCommercial–NoDerivatives (CC BY-NC-ND 4.0) license. ✅ Free for research and non-commercial use ❌ Commercial use and derivative redistribution are not permitted 🖼️ Source images follow the Open Images (CC BY 2.0) license By using this dataset, you agree to comply with the terms of both licenses.

If you use 🍌 Pico-Banana-400K in your research, please cite it as follows:

@inproceedings{Qian2025PicoBanana400KAL,
  title={Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing},
  author={Yusu Qian and Eli Bocek-Rivele and Liangchen Song and Jialing Tong and Yinfei Yang and Jiasen Lu and Wenze Hu and Zhe Gan},
  year={2025},
  url={https://api.semanticscholar.org/CorpusID:282272484}
}