MICo-150K: 150,000 Synthetic Images for Multi-Image Composition

MICo-150K consists of 150,000 synthetic image records generated by Nano Banana and Nano Banana Pro for multi-image composition (MICo) research. Created by kr-cen and updated in February 2026, the dataset supports the development of open-source vision-language models for complex image-to-image tasks.

Use Cases

• Fine-tuning models for human-centric tasks to improve realistic human placement in generated scenes
• Training image-to-image models on decomposition and recomposition tasks to isolate specific objects
• Developing multi-image composition workflows using the human-object interaction labels

Strengths

• 150,000 synthetic records
• Apache 2.0 license
• Covers four distinct MICo task categories including human-object interaction

Limitations

• Synthetic data may contain visual artifacts not present in natural photography
• Lack of documented schema for specific metadata columns in the primary description

Provenance

Source

kr-cen via Hugging Face (Arxiv 2512.07348)

Collection Method

synthetic

Freshness

Last updated February 2026.