Introduction
Text-guided image editing is a core capability for modern multimodal large language models (MLLMs). However, the field faces a significant data bottleneck. Existing datasets are often either small and human-curated, which limits scalability, or entirely synthetic, leading to a "domain shift" where models struggle with real-world photographs. Furthermore, the lack of "negative" examples—failed edits—hinders the application of alignment techniques like Direct Preference Optimization (DPO).
To address these challenges, researchers have introduced Pico-Banana-400K, a massive dataset of approximately 400,000 high-quality text-guided image edits derived from real-world photographs. This resource includes single-turn instructions, multi-turn editing sequences, and preference pairs for alignment research.
Methodology
The creation of Pico-Banana-400K utilizes a systematic pipeline designed to scale data generation while maintaining high quality through automated oversight.
The pipeline consists of four main stages:
- Source Material Selection: Images are sourced from the OpenImages collection to ensure visual realism.
- Instruction Generation: Gemini-2.5-Flash generates both detailed prompts for training and concise user-style commands.
- Execution: The Nano-Banana model performs the edits based on these instructions.
- Automated Quality Control: Gemini-2.5-Pro acts as a judge, scoring results based on instruction compliance (40%), seamlessness (25%), preservation of original content (20%), and technical quality (15%). Failed edits are retried up to three times; persistent failures are retained as negative samples for preference learning.
Dataset Characteristics
The resulting dataset comprises 386,000 samples categorized into three distinct subsets to support various training objectives.
- Single-Turn SFT (258K): High-quality pairs for Supervised Fine-Tuning.
- Multi-Turn SFT (72K): Sequences of 2–5 edits that require the model to maintain context and handle referential language (e.g., "change the color of it").
- Preference Pairs (56K): Successful vs. failed edit pairs for reward modeling and DPO.
The dataset spans 35 edit types across 8 categories, including stylistic transfers, object manipulation, and photometric changes.
Results and Analysis
Analysis of the dataset reveals varying levels of difficulty for current MLLMs:
- High Success Rates: Global style transfers (93.4%) and vintage filters (90.7%) are handled with high reliability.
- Moderate Success: Object removal is robust, succeeding in 83.3% of cases, though it can occasionally affect adjacent regions.
- Current Frontiers: Tasks requiring precise spatial reasoning or typography remain challenging. Relocating objects (59.2%) and changing font styles (57.6%) represent the most difficult tasks for current models.
Conclusion and Future Work
Pico-Banana-400K provides a robust foundation for training the next generation of image-editing models by bridging the gap between synthetic scale and real-world realism. While it significantly advances the availability of training data, limitations remain in spatial precision, typography integrity, and identity preservation during transformations. Future work should focus on region-referential prompting and explicit supervision for text rendering to overcome these hurdles.
Summary updated