Apple Releases Pico-Banana-400K Dataset for Text-Guided Image Editing

Apple Releases Pico-Banana-400K Dataset for Text-Guided Image Editing

Apple researchers have introduced Pico-Banana-400K, a large-scale dataset of 400,000 images designed to facilitate the development of text-guided image editing models. This dataset addresses a critical gap in the availability of high-quality, shareable image editing data by combining real-world photographs with systematic editing processes and rigorous quality filtering. The dataset is publicly accessible under a Creative Commons license, aiming to democratize research in this domain.

---

Dataset Overview

• Size: 400,000 images, with 257,000 generated via single-turn prompts and 56,000 retained as failure cases.
• Purpose: Enable training of text-guided image editing models by providing diverse, high-quality, and curated examples.
• Key Differentiators:
  • Systematic quality control via MLLM-based scoring (Gemini-2.5-Pro).
  • Focus on instruction faithfulness (alignment between text prompts and visual edits).
  • Coverage of 35 edit types across eight categories, including object-level semantics and stylistic transformations.

---

Creation Process

1. Source Data:

   • Real photographs from the Open Images collection (licensed under CC BY 2.0).
   • Images include humans, objects, and textual scenes.

2. Editing Pipeline:

   • Nano-Banana (Google’s model) was used to apply text-guided edits to the source images.
   • Gemini-2.5-Pro filtered results based on four criteria:
     • Instruction compliance (40% weight): Ensures edits match the prompt.
     • Editing realism (25%): Measures how natural the edits appear.
     • Preservation balance (20%): Maintains original content integrity.
     • Technical quality (15%): Evaluates resolution, artifacts, and other technical aspects.

3. Prompt Generation:

   • Gemini-2.5-Flash generated initial prompts, which were then condensed into human-like instructions using Qwen2.5-7B-Instruct.

---

Specialized Subsets

The dataset includes three subsets tailored for specific research goals:

• Multi-Turn Instructions (72K examples):
  • Designed for sequential editing tasks, enabling research into reasoning and planning across multiple edits.
• Failed Edits (56K examples):
  • Retained for robustness training and reward model development.
• Instruction Rewriting Subset:
  • Pairs long and short prompts to aid in instruction summarization and rewriting capabilities.

---

Licensing and Accessibility

• Pico-Banana-400K: Available on Apple’s CDN via GitHub under CC BY-NC-ND 4.0 (non-commercial, no derivatives).
• Open Images Originals: Licensed under CC BY 2.0 for research and commercial use.

---

Impact and Use Cases

• Research Advancement: Addresses limitations of existing datasets (e.g., small size or proprietary synthetic data).
• Model Training: Supports alignment research, reward modeling, and instruction-following capabilities in AI systems.
• Practical Applications: Potential use in tools for photo editing, content creation, and AI-assisted design workflows.

---

Reference

https://www.infoq.com/news/2025/11/apple-pico-banana-image-editing/