BryceJia/MICo-150K

--- license: apache-2.0 tags: - Multi-Image Composition - Image Editing - Image Generation size_categories: - 100K<n<1M task_categories: - image-to-image language: - en --- # MICo-150K Dataset [](https://www.arxiv.org/pdf/2512.07348) [](https://www.arxiv.org/abs/2512.07348) [](https://github.com/A113N-W3I/MICo-150K) [](https://mico-150k.github.io/) ## 🌟 Catalogue 1. [Dataset Details](#🔍-dataset-details) 2. [Data Structure](#📃-data-structure) 3. [Gallery: Decompose & Recopmose](#gallery-dere-subset) 4. [Gallery: Human Centric Tasks](#gallery-human-centric) 5. [Gallery: Object Centric Tasks](#gallery-object-centric) 6. [Gallery: Human Object Interaction](#gallery-hoi) ## 📖 Introduction **MICo-150K** is a large-scale synthetic dataset generated by **Nano Banana** and **Nano Banana Pro**, designed to advance open-source models in **M**ulti-**I**mage **Co**mposition (MICo). We fine-tune a diverse set of base models—including [Qwen-Image](https://github.com/QwenLM/Qwen-Image), [BAGEL](https://bagel-ai.org/), [OmniGen2](https://github.com/VectorSpaceLab/OmniGen2), [Lumina-DiMOO](https://github.com/Alpha-VLLM/Lumina-DiMOO), and [BLIP3o-Next-Edit](https://github.com/JiuhaiChen/BLIP3o), on MICo-150K. All models demonstrate substantial performance improvements on our proposed MICo Bench after fine-tuning. Notably, Qwen-Image, originally developed as a text-to-image model, achieves remarkable gains after adaptation. The fine-tuned variant, Qwen-Image-MICo, surpasses Qwen-Image-Edit-2509 on both MICo Bench and OmniContext benchmarks, highlighting its strong generalization capability and broad applicability in multi-image composition scenarios.  ## 🔍 Dataset Details We organize the MICo-150K dataset into **three primary categories** and **one specialized task subset, [De&Re](#gallery-dere-subset)**. Each primary category encompasses multiple sub-tasks: * [Human-Centric Tasks](#gallery-human-centric) * [Object-Centric Tasks](#gallery-object-centric) * [Human–Object Interaction (HOI) Tasks](#gallery-hoi) ### Human-Centric Tasks 1. Two Persons (3K samples each for two males, two females, and one male–one female scenarios) 2. Three Persons (3K samples each for three males, three females, one male–two females, and two males–one female scenarios) 3. Person(s) + Scene (6K samples for one person + scene and 6K samples for two persons + scene) ### Object-Centric Tasks 1. Multi-Object Composition (5K samples each for two, three, four, and five objects) 2. Object(s) + Scene (5K samples each for one object + scene and two objects + scene) ### Human–Object Interaction (HOI) Tasks 1. Person + Apparel (6K samples each for one person with one, two, three, and four apparel items) 2. Person + Object (6K samples each for one person + one object, one person + two objects, two persons + one object, and two persons + two objects) 3. Person + Apparel + Object (6K samples each for 1H1C1O, 1H1C2O, 1H2C1O, and 1H2C2O configurations) ### De&Re Subset The De&Re subset focuses on decomposition and recomposition tasks. Specifically, elements from a real-world image (e.g., persons, apparel, scenes) are first decomposed into multiple component images. These components are subsequently recomposed into a single composite image according to a specified instruction. This subset contains 11K samples in total. ## 📃 Data Structure For Human-Centric, Object-Centric, and HOI tasks, the Parquet files share a unified schema with the following keys: * `input_images`: a list of input images * `output_image`: the composed image * `instruction`: a descriptive or imperative instruction specifying how to compose the input images * `separate_prompt`: captions corresponding to each input image * `editing_type`: a label describing the task type For the De&Re subset, the Parquet files contain the following keys: * `reference`: the real image used for decomposition * `input`: a list of images obtained by decomposing the `reference` image * `output`: the recomposed image * `instruction`: a description of how to compose the decomposed input images * `separate_prompt`: captions corresponding to each decomposed input image * `editing_type`: a label describing the task type ## 🖼️ Gallery ### Gallery: De&Re Subset       ### Gallery: Human Centric     ### Gallery: Object Centric    ### Gallery: HOI     ## ✨ Citation If you find this dataset or the associated work useful for your research, please cite the paper: ```bib @article{wei2025mico, title={MICo-150K: A Comprehensive Dataset Advancing Multi-Image Composition}, author={Wei, Xinyu and Cen, Kangrui and Wei, Hongyang and Guo, Zhen and Li, Bairui and Wang, Zeqing and Zhang, Jinrui and Zhang, Lei}, journal={arXiv preprint arXiv:2512.07348}, year={2025} } ```license: apache-2.0 tags: - 多图像合成（Multi-Image Composition） - 图像编辑（Image Editing） - 图像生成（Image Generation） size_categories: - 100K < n < 1M task_categories: - 图像到图像 language: - 英语 # MICo-150K 数据集 [](https://www.arxiv.org/pdf/2512.07348) [](https://www.arxiv.org/abs/2512.07348) [](https://github.com/A113N-W3I/MICo-150K) [](https://mico-150k.github.io/) ## 🌟 目录 1. [数据集详情](#🔍-数据集详情) 2. [数据结构](#📃-数据结构) 3. [图集：分解与重组](#gallery-dere-subset) 4. [图集：以人为中心的任务](#gallery-human-centric) 5. [图集：以物体为中心的任务](#gallery-object-centric) 6. [图集：人-物体交互任务](#gallery-hoi) ## 📖 简介 **MICo-150K** 是由**Nano Banana**与**Nano Banana Pro**生成的大规模合成数据集，旨在推动开源模型在**多图像合成（Multi-Image Composition，MICo）**领域的发展。 我们针对MICo-150K数据集微调了多款基础模型，包括[Qwen-Image](https://github.com/QwenLM/Qwen-Image)、[BAGEL](https://bagel-ai.org/)、[OmniGen2](https://github.com/VectorSpaceLab/OmniGen2)、[Lumina-DiMOO](https://github.com/Alpha-VLLM/Lumina-DiMOO)以及[BLIP3o-Next-Edit](https://github.com/JiuhaiChen/BLIP3o)。所有模型在我们提出的MICo基准测试（MICo Bench）上经过微调后，性能均有显著提升。值得注意的是，原本作为文本到图像模型开发的Qwen-Image，在适配后取得了亮眼的性能提升。其微调变体Qwen-Image-MICo在MICo基准测试与OmniContext基准测试中均超越了Qwen-Image-Edit-2509，展现出极强的泛化能力与在多图像合成场景中的广泛适用性。  ## 🔍 数据集详情 我们将MICo-150K数据集划分为**三大核心类别**与**一个专属任务子集——[De&Re](#gallery-dere-subset)**。每个核心类别包含多项子任务： * [以人为中心的任务](#gallery-human-centric) * [以物体为中心的任务](#gallery-object-centric) * [人-物体交互（Human–Object Interaction，HOI）任务](#gallery-hoi) ### 以人为中心的任务 1. 双人场景：针对两名男性、两名女性以及一男一女三种组合，每种组合各含3000个样本 2. 三人场景：针对三名男性、三名女性、一男两女以及两男一女四种组合，每种组合各含3000个样本 3. 人物+场景：单人+场景与双人+场景各含6000个样本 ### 以物体为中心的任务 1. 多物体合成：针对2、3、4、5个物体的组合，每种组合各含5000个样本 2. 物体+场景：单个物体+场景与两个物体+场景各含5000个样本 ### 人-物体交互（HOI）任务 1. 人物+服饰：针对一人搭配1、2、3、4件服饰的场景，每种场景各含6000个样本 2. 人物+物体：针对一人+一件物体、一人+两件物体、两人+一件物体以及两人+两件物体四种组合，每种组合各含6000个样本 3. 人物+服饰+物体：针对1H1C1O、1H1C2O、1H2C1O以及1H2C2O四种配置，每种配置各含6000个样本 ### De&Re子集 该子集聚焦于分解与重组任务。具体而言，首先将现实图像中的元素（如人物、服饰、场景）分解为多张组件图像，随后根据指定指令将这些组件重组为一张合成图像。该子集总计包含11000个样本。 ## 📃 数据结构 对于以人为中心、以物体为中心以及人-物体交互任务，Parquet文件采用统一的Schema，包含以下字段： * `input_images`：输入图像列表 * `output_image`：合成后的图像 * `instruction`：用于指定输入图像合成方式的描述性或指令性文本 * `separate_prompt`：与每张输入图像对应的描述性标题 * `editing_type`：描述任务类型的标签 对于De&Re子集，Parquet文件包含以下字段： * `reference`：用于分解的原始现实图像 * `input`：通过分解`reference`图像得到的组件图像列表 * `output`：重组后的合成图像 * `instruction`：用于描述如何合成分解得到的输入图像的文本 * `separate_prompt`：与每张分解得到的输入图像对应的描述性标题 * `editing_type`：描述任务类型的标签 ## 🖼️ 图集 ### 图集：De&Re子集       ### 图集：以人为中心的任务     ### 图集：以物体为中心的任务    ### 图集：人-物体交互任务     ## ✨ 引用 如果您的研究中用到了本数据集或相关工作，请引用该论文： bib @article{wei2025mico, title={MICo-150K: A Comprehensive Dataset Advancing Multi-Image Composition}, author={Wei, Xinyu and Cen, Kangrui and Wei, Hongyang and Guo, Zhen and Li, Bairui and Wang, Zeqing and Zhang, Jinrui and Zhang, Lei}, journal={arXiv preprint arXiv:2512.07348}, year={2025} }