🍌 Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing

Author: Apple AIML Research | Published: 2025-10-21 | Generated: 2025-10-26 | Domain: github.com
Tags: ‘#image-editing’ ‘#multimodal-ai’ ‘#dataset’ ‘#computer-vision’ ‘#preference-learning’ ‘#gemini’

---

TLDR

Pico-Banana-400K is a roughly 400K-example dataset for text-guided image editing, spanning 35 edit operations in eight semantic categories and supporting single-turn SFT, preference learning, and multi-turn editing. Instructions are generated from Open Images inputs with Gemini-2.5-Flash; Nano-Banana produces edits, while Gemini-2.5-Pro scores instruction compliance, realism, preservation, and technical quality. Successful edits exceeding an approximately 0.7 threshold form the main training sets, while about 56K failures are retained for robustness and preference modeling.

Key Takeaways

• Dataset composition: Includes approximately 257K successful single-turn SFT triplets, 56K preference-learning examples with positive and negative images, and 72K successful multi-turn samples.
• Broad edit taxonomy: Covers 35 operations across eight categories: pixel/photometric, object-level, scene composition, stylistic, text/symbol, human-centric, scale/perspective, and spatial/layout edits.
• Quality-control pipeline: Gemini-2.5-Pro evaluates edits using weighted criteria: instruction compliance (40%), editing realism (25%), preservation balance (20%), and technical quality (15%).
• Distribution is object-centric: Object-level semantic edits account for 35% of the data; scene composition/multi-subject edits account for 20%, and human-centric edits 18%.
• Access and licensing: Edited-image manifests and JSONL metadata are hosted on Apple’s CDN; source Open Images files must be downloaded separately. The dataset uses CC BY-NC-ND 4.0, prohibiting commercial use and derivative redistribution.

Images & Media

• None present in article.

Referenced Links

• Pico-Banana-400K paper (arXiv PDF) — Research paper describing the dataset.
• Open Images — Source image dataset used for original images.
• Gemini 2.5 Flash — Model used to generate image-grounded editing instructions.
• Single-turn SFT manifest — Download manifest for successful single-turn edited images.
• Preference-data manifest — Download manifest for preference-learning samples.
• Multi-turn manifest — Download manifest for multi-turn edited images.
• Open Images download instructions — Instructions for obtaining source-image archives.
• URL-to-local-path mapping script — Utility for mapping Open Images URLs to locally downloaded files.
• Pico-Banana-400K arXiv record — Citation and paper metadata.