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

TL;DR

Pico-Banana-400K tackles the lack of large, open, real-image editing datasets by presenting a ~$4\times 10^{5}$-example corpus built from OpenImages using Nano-Banana and validated with Gemini-2.5-Pro, organized into $35$ edit types across $8$ categories and augmented with $3$ specialized subsets. The approach includes dual instruction generation (detailed and concise), automated quality scoring, and a multi-turn extension (72K sequences) to study iterative editing and alignment with $56{,}$000 preference pairs for reward-model research. With rigorous per-edit-type analyses showing strong performance on global and stylistic edits but remaining challenges in precise geometry and typography, the dataset provides a robust platform for training and benchmarking next-generation text-guided image editing models. The work enables scalable, high-quality instruction-faithful data and sets the stage for extensive benchmarking, model-training studies, and exploration of controllability and fidelity in real-image editing systems.

Abstract

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.

Figures (6)

• Figure 1: Pico-Banana-400K dataset overview. The pipeline (bottom) shows how diverse OpenImages inputs are edited using Nano-Banana and quality-filtered by Gemini-2.5-Pro, with failed attempts automatically retried. The dataset comprises 386K examples across single-turn SFT (66.8%), preference pairs (14.5%), and multi-turn sequences (18.7%), organized by our comprehensive edit taxonomy (top left).
• Figure 2: Example single-turn text-guided image edits from the Pico-Banana-400K dataset. Each pair shows the edited result (right) and its corresponding original image (left). The dataset spans diverse edit types, including photometric adjustments, object-level manipulations, stylistic transformations, and scene or lighting modifications. These examples illustrate the visual diversity, realism, and high instruction fidelity achieved by the Nano-Banana editing model.
• Figure 3: Distribution of image editing instruction content.
• Figure 4: Preference triplet example. From left to right: the original image, the natural-language instruction (center panel) requesting relocation of the pink--white straws into the leftmost glasses, and two model outputs: a successful edit that satisfies the instruction and preserves scene context, and a failed edit that violates the instruction (incorrect placement/geometry). Such (success, failure) pairs are retained as preference data for alignment studies.
• Figure 5: Multi-turn image editing example. Starting from the original pumpkin image, the model first applies a vintage film grain effect, replaces the dark background with a haunted house scene, transforms the entire setting into a snowy winter landscape, and finally adjusts the global lighting to a warm, golden-hour glow, producing the final image on the right.