New MICo-150K dataset and benchmark advance multi-image composition tasks

Researchers have introduced MICo-150K, a large-scale dataset designed to improve multi-image composition (MICo) capabilities in AI models. The dataset addresses the challenge of synthesizing coherent images from multiple references by categorizing MICo into seven tasks and providing high-quality composite images. MICo-150K includes a unique subset for real-world image decomposition and recomposition, along with a benchmark suite and a new evaluation metric called Weighted-Ref-VIEScore. Fine-tuning models on this dataset has shown significant improvements in MICo tasks, with a baseline model, Qwen-MICo, demonstrating enhanced performance. AI

影响 Enhances AI's ability to generate complex images from multiple references, potentially improving creative tools and visual content generation.

排序理由 The cluster describes a new academic paper introducing a dataset and benchmark for multi-image composition.

在 arXiv cs.CV 阅读 →

• Kangrui Cen
• MICo-150K
• MICo-Bench
• Qwen-Image-2509
• Qwen-Image-Edit
• Qwen-MICo
• Weighted-Ref-VIEScore

• 论文
• 其他

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →