strong-followup
Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing
Yusu Qian, Eli Bocek-Rivele, Liangchen Song · Apple · 2025-10 · arXiv:2510.19808
Why it matters
400K edit pairs on real photographs, distilled from a frontier closed editor onto OpenImages. Notable both as Apple entering open editing data and as evidence that closed-model distillation is now a standard way to build open training sets.
What this paper does
read: full textBefore this
The strongest instruction editors, such as GPT-4o and Nano-Banana, are closed systems. Openly available training sets for instruction editing were comparatively small, built with domain shift from real photographs, uneven coverage across edit types, and inconsistent quality control, since most relied on proprietary generators or limited human curation.
The problem
There was no large-scale, openly released, systematically quality-controlled dataset of instruction edits grounded in real photographs, and no open resource supporting multi-turn editing, preference-based alignment, or instruction rewriting research together.
The idea
Distill edit pairs from a frontier closed editor onto real photographs at scale, organize them under an explicit edit taxonomy, and filter them with an automated multimodal judge, releasing the result openly along with subsets built for multi-turn, preference, and instruction-length research.
How it works
This is a dataset, not an editing method. Source images are real photographs drawn from OpenImages. Nano-Banana, Google's Gemini-2.5-Flash-Image, generates edit pairs across 35 edit types grouped into 8 categories, namely pixel and photometric, object-level semantic, scene composition, stylistic, text and symbol, human-centric, scale, and spatial or layout edits. Each pair is given two instructions, a detailed training-oriented version written by Gemini-2.5-Flash and a short user-style version produced by rewriting it with Qwen2.5-7B. Quality control is automated, with Gemini-2.5-Pro acting as a judge scoring each pair on four weighted criteria, instruction compliance at 40%, seamlessness at 25%, preservation balance at 20% and technical quality at 15%, and pairs are kept above a threshold of roughly 0.7. Beyond the main single-turn set, the release includes a 72K-example multi-turn subset of 2 to 5 sequential edits per image, a 56K-example preference subset pairing successful against failed edits for reward modeling and DPO-style alignment work, and paired long-short instructions across the dataset for instruction rewriting and summarization research.
Evidence
The release totals roughly 400K examples, split into 258K single-turn edits that passed quality filtering (a 66.8% success rate among generated attempts), 56K preference pairs, and 72K multi-turn sequences. Production cost was approximately $100K. Success rate varies sharply by edit type, from 93.4% for style transfer and 90.7% for film grain to 59.2% for object relocation and 57.6% for font changes.
Limitations
The paper concedes that the underlying generator itself, not just the filtering pipeline, struggles with fine-grained spatial editing, layout extrapolation, and typography. Geometry-heavy edits show perspective inconsistencies and topology breaks, and text edits are brittle on letterform integrity, alignment and contrast. Because every pair is distilled from one closed model's outputs rather than independently verified, the dataset inherits that model's specific failure modes and biases wherever the automated judge does not catch them.
Why it matters
It establishes closed-model distillation onto real photographs, filtered by an automated multimodal judge, as a reproducible way to build large open instruction-editing datasets, and it is notable as Apple's first open release of editing training data. The preference and multi-turn subsets extend what open data can support beyond single-shot supervised fine-tuning.
Abstract, in the authors' own words
Recent advances in multimodal models have demonstrated remarkable text-guided image editing capabilities, with systems like GPT-4o and Nano-Banana setting new benchmarks. However, the research community's progress remains constrained by the absence of large-scale, high-quality, and openly accessible datasets built from real images. We introduce Pico-Banana-400K, a comprehensive 400K-image dataset for instruction-based image editing. Our dataset is constructed by leveraging Nano-Banana to generate diverse edit pairs from real photographs in the OpenImages collection. What distinguishes Pico-Banana-400K from previous synthetic datasets is our systematic approach to quality and diversity. We employ a fine-grained image editing taxonomy to ensure comprehensive coverage of edit types while maintaining precise content preservation and instruction faithfulness through MLLM-based quality scoring and careful curation. Beyond single turn editing, Pico-Banana-400K enables research into complex editing scenarios. The dataset includes three specialized subsets: (1) a 72K-example multi-turn collection for studying sequential editing, reasoning, and planning across consecutive modifications; (2) a 56K-example preference subset for alignment research and reward model training; and (3) paired long-short editing instructions for developing instruction rewriting and summarization capabilities. By providing this large-scale, high-quality, and task-rich resource, Pico-Banana-400K establishes a robust foundation for training and benchmarking the next generation of text-guided image editing models.
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.
This is a strong-followup entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.