Apple releases 400K image dataset to improve AI editing
Apple researchers have released Pico-Banana-400K, a comprehensive dataset containing 400,000 curated images designed to improve how artificial intelligence systems edit photos based on text prompts, the company announced in a research paper published this week.
The massive dataset aims to address what Apple describes as a critical gap in current AI image editing training, where progress has been constrained by inadequate datasets built from real photographs. While systems like GPT-4o can make impressive edits, researchers say the lack of large-scale, high-quality training data has limited advancement in the field.
Summaries 3 keypoint:
- Apple released Pico-Banana-400K, a 400,000-image, non-commercial research dataset for text-guided image editing, built from real photos and curated to advance training and benchmarking of editing models.
- The dataset emphasizes quality and diversity via a taxonomy of 35 edit types across 8 categories, automated judging with Gemini-2.5-Pro, and three subsets: 258K single-edit pairs, 56K preference pairs, and 72K multi-turn sequences.
- Findings highlight current model limits: global style edits perform strongly, but fine-grained tasks like object relocation and text editing remain below ~60% success, underscoring areas for improvement in spatial and typographic control.
Systematic Approach to Quality and Diversity
What distinguishes Pico-Banana-400K from previous datasets is Apple's systematic approach to quality control and comprehensive coverage. The images are organized into 35 different edit types across eight categories, ranging from basic adjustments like color changes to complex transformations such as converting people into Pixar-style characters or LEGO figures.[1][2][3][4]
Apple built the dataset using Google's Gemini-2.5-Flash-Image model, also known as Nano-Banana, to generate the edits, while Gemini-2.5-Pro served as an automated quality control system to evaluate results based on instruction compliance and technical quality. Each image in the set underwent this rigorous AI-powered screening process before inclusion.[3][4][5]
The dataset includes three specialized subsets: 258,000 single-edit examples for basic training, 56,000 preference pairs comparing successful and failed edits, and 72,000 multi-turn sequences showing how images evolve through multiple consecutive edits.[2][4][3]
Revealing Current AI Limitations
Apple's research revealed significant limitations in current image editing models. While global style changes succeeded 93% of the time, precise tasks like relocating objects or editing text struggled with success rates below 60%. These findings provide valuable insights into where AI image editing still falls short of user expectations.[1][2]
The complete Pico-Banana-400K dataset is freely available for non-commercial research use on GitHub, allowing developers and researchers to use it for training more capable image editing AI systems. According to the researchers, the dataset establishes "a robust foundation for training and benchmarking the next generation of text-guided image editing models".[2][3][1]
Reference source
[1](https://www.macrumors.com/2025/10/29/apple-ai-dataset-improve-photo-editing-models/)
[2](https://9to5mac.com/2025/10/28/apple-pico-banana-400k-dataset/)
[3](https://www.inkl.com/news/apple-releases-new-ai-dataset-to-improve-image-editing-models)
[1](https://appleinsider.com/articles/25/10/29/apple-wants-to-improve-everybodys-ai-image-editors-with-new-training-dataset)
[2](https://www.inkl.com/news/apple-releases-new-ai-dataset-to-improve-image-editing-models)
[3](https://www.macrumors.com/2025/10/29/apple-ai-dataset-improve-photo-editing-models/)
[4](https://9to5mac.com/2025/10/28/apple-pico-banana-400k-dataset/)
[5](https://9to5mac.com/2025/10/28/apple-pico-banana-400k-dataset/?extended-comments=1)