Source-linked AI summary
Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing
Yusu Qian, Eli Bocek-Rivele, Liangchen Song, Jialing Tong, Yinfei Yang, Jiasen Lu, Wenze Hu, Zhe Gan
TL;DR
Open research lacks large-scale, high-quality, shareable editing datasets built from real images. Pico-Banana-400K addresses this gap with a systematically constructed, quality-controlled dataset, while its analysis identifies global transformations as more reliable than fine-grained spatial editing and typography.
Problem
Open research lacks large-scale, high-quality, fully shareable editing datasets built from real images, with existing resources often showing domain shifts, imbalanced edit types, and inconsistent quality control.
Method
The dataset uses real OpenImages photographs, Nano-Banana edits across 35 edit types, automated multi-dimensional Gemini-2.5-Pro scoring, manual verification, and preserved failed attempts as negative examples.
Results
Pico-Banana-400K provides approximately 400K high-quality, publicly released image-editing examples organized by a 35-type taxonomy, with global stylistic transformations generally more reliable than precise geometry, layout, and typography edits.
Takeaways & Limitations
The dataset provides a shareable resource for open text-guided image-editing research and highlights fine-grained spatial editing, layout extrapolation, and typography as open problems.
Takeaways & Limitations
Under text-only conditioning, edits can localize imperfectly and cause incidental nearby changes or modest color and texture drift.
Abstract
from arXiv · showhide
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.
1 Introduction
Pico-Banana-400K addresses the shortage of large-scale, high-quality, shareable editing datasets by combining real OpenImages photographs with systematic generation, taxonomy coverage, and quality control. It also supports single-turn, preference, multi-turn, and diverse instruction-format research.
- Existing editing datasets often depend on proprietary synthetic models or small human-curated subsets, with domain shifts, imbalanced edit types, and inconsistent quality control.
- Pico-Banana-400K provides approximately 400K text-guided image edits built from real OpenImages photographs under clear licensing terms.
- The construction pipeline uses Nano-Banana to generate edits across 35 edit types and Gemini-2.5-Pro to score instruction compliance, editing quality, preservation balance, and technical quality.
- Failed editing attempts are automatically retried, while successful edits form the core training data and failed edits are preserved as negative examples.
- The dataset includes approximately 400K high-quality examples organized by a 35-type taxonomy with automated scoring and manual verification.
- Its research resources include 56K preference pairs for alignment and reward modeling, plus 72K multi-turn sequences for iterative editing, context-aware editing, and planning.
2 Dataset Construction
Pico-Banana-400K is constructed from OpenImages using a 35-type editing taxonomy, dual instruction formats, automated judging, retries, and specialized data products for single- and multi-turn editing.
- Source Images and Taxonomy: The dataset uses OpenImages images covering humans, objects, and textual scenes, organized into 35 edit types across eight categories.Each image-instruction pair receives one primary edit type, with category-specific filtering for human-centric and text-related operations.
- Source Images and Taxonomy: Three unreliable operation groups are excluded: weak photometric changes, strong perspective or pose rewrites, and two-image compositions.These operations produced negligible or unstable changes, structural artifacts, or insufficiently reliable training pairs.
- Dual Instruction Generation: Each edit has complementary long, detailed and concise, human-style instruction variants for richly specified supervision and natural user prompts.Gemini-2.5-Flash generates detailed prompts, while Qwen produces concise rewrites using human-written demonstrations.
- Quality Assessment: Nano-Banana executes edits and Gemini-2.5-Pro judges instruction compliance, seamlessness, preservation balance, and technical quality before retention.The resulting quality score uses a strict threshold of approximately 0.7 to classify successful and failed edits.
- Single-Turn Dataset: Approximately 258K successful edits form the main dataset, while approximately 56K failures are retained as paired negative examples for preference learning.The table procedure allows up to three retries; failures after all attempts are discarded, while earlier failed attempts can be saved with a later success.
3 Dataset Analysis
Dataset analysis shows that global appearance and style edits are most reliable, whereas precise spatial, layout, typography, and some human-stylization edits remain challenging.
- Easy: global edits and stylization: Global edits achieve the highest reliability, including strong artistic style transfer at 0.9340 and film grain/vintage at 0.9068.These operations mainly reshape global texture, color statistics, and tone, requiring limited spatial reasoning or object coordination.
- Moderate: object semantics and scene context: Remove object reaches 0.8328 and replace category 0.8348, while seasonal change reaches 0.8015 and photo→cartoon/sketch 0.8006.Object-semantic and scene-context edits are generally robust but can suffer from imperfect localization and color or texture drift.
- Hard: precise geometry, layout, and typography: Change font/style has the lowest rate at 0.5759, with other text operations remaining unstable because of letterform, alignment, and contrast challenges.Human stylizations also show identity drift and shading artifacts, including caricature at 0.5884.
- Implications: The analysis identifies stronger spatial conditioning, geometry-aware objectives, text-rendering supervision, and identity-preserving constraints as promising directions.These directions target the observed weaknesses in spatial editing, layout extrapolation, typography, and human-centric stylization.
4 Related Work
Related work spans real-edit datasets, synthetic or mixed pipelines, training-free editors, and finetuning-based models. Pico-Banana-400K is positioned around quality control, instruction fidelity, fine-grained coverage, and specialized subsets rather than sheer scale.
- Image Editing Datasets: GIER and MagicBrush collect grounded real-image edits, while later datasets expand category coverage, masks, and visual diversity through synthetic or mixed pipelines.MagicBrush includes 10K human-annotated triplets spanning single- and multi-turn edits.
- Image Editing Datasets: Table 3 provides a side-by-side comparison of representative image editing datasets.The supplied caption identifies the table’s comparative purpose but does not specify individual comparison fields or outcomes.
- Image Editing Datasets: Recent datasets distill frontier multimodal models, including Echo-4o-Image with ∼180K synthetic examples and GPT-Image-Edit-1.5M with 1.5M regenerated triplets.Pico-Banana-400K also belongs to this distillation trend, using Nano-Banana.
- Image Editing Models: Training-free methods use noising–denoising trajectories, attention manipulation, or cross-attention control, whereas finetuning-based methods learn from instruction, before, and after triplets.The cited families include SDEdit, Prompt-to-Prompt, DiffEdit, and InstructPix2Pix.
- Positioning: Pico-Banana-400K emphasizes quality-controlled, instruction-faithful edits and fine-grained category coverage rather than sheer scale.Its positioning also includes a 56K preference-triplet subset and human-centric realistic and stylized transformations.
5 Conclusion
The conclusion presents Pico-Banana-400K as a publicly released resource and scalable framework for text-guided image-editing research. It combines instruction generation, image editing, and verification while leaving benchmarking and training studies for future work.
- 5 Conclusion: Pico-Banana-400K combines Gemini-2.5-Flash instruction generation, Nano-Banana editing, and Gemini-2.5-Pro verification in a scalable framework.The dataset is aimed at advancing image-editing research.
- 5 Conclusion: All images and metadata are publicly released to support open research in text-guided image editing.The conclusion identifies model benchmarking and training studies as future work.
A System Prompt for Edit Instruction Generation
The instruction-generation prompt directs Gemini-2.5-Flash to write concise, natural, photorealistic editing commands grounded in visible image content. It specifies a JSON output containing a prompts array.
- System prompt: Gemini-2.5-Flash is instructed to act as a professional photo-editing prompt writer and produce natural instructions reflecting plausible user intents.The system prompt requests one concise natural-language instruction for an image-editing model.
- System prompt: Generated instructions must be grounded in visible objects, colors, positions, and closely related to the image content.This constraint ties each prompt to observable image information.
- Output format: The output format is a JSON object with a prompts array containing photorealistic prompts.The supplied example shows the prompts array structure.
B System Prompt used by Gemini-2.5-Pro as a Judge
Gemini-2.5-Pro evaluates edited images against their inputs and instructions using weighted criteria that measure compliance, seamlessness, preservation, and technical quality.
- The judge receives the edited result, reference input images, and editing instruction, then uses all three for assessment.
- Edit Instruction Compliance carries the largest weight at 40% and measures whether requested changes are visible, properly implemented, and aligned with the intended edit.
- Editing Quality & Seamlessness receives 25% and evaluates realism, artifacts, blending, lighting, and perspective preservation.
- Preservation vs. Change Balance accounts for 20% by checking that appropriate original elements remain while unrelated regions stay focused and unaffected.
- Technical Quality contributes 15% based on sharpness, color consistency, exposure, and the absence of artifacts or distortions.
- The evaluation compares the edited result with the original image to verify successful changes, preserved elements, and accurate instruction interpretation.