MICo-150K: A Comprehensive Dataset Advancing Multi-Image Composition

Official repository for the paper MICo-150K: A Comprehensive Dataset for Multi-Image Composition.

📢 News

• Apr 15, 2026: 📊 Release MICo-Bench — 897 evaluation cases with Weighted-Ref-VIEScore evaluation script. See MICo-Bench/.
• Mar 1, 2026: 🔥 Release Qwen-Image-MICo checkpoint and inference script.
• Feb 21, 2026: 🎉 MICo-150K has been accepted to CVPR 2026!
• Feb 21, 2026: 📦 We released the full MICo-150K dataset on Hugging Face: https://huggingface.co/datasets/kr-cen/MICo-150K.
• Dec 16, 2025: 🔥 We released official gradio demo for Qwen-Image-MICo, try it out!
• Dec 10, 2025: 🚀 We released finetuned checkpoints BAGEL-MICo, BLIP3o-Next-MICo, Lumina-DiMOO-MICo, and OmniGen2-MICo, with impressive multi-image composition capability. Our MICo-150K dataset coming soon, stay tuned! 👀
• Dec 10, 2025: 📖 We released multi-image composition training & inference guideline for community models. Our finetuned checkpoints coming soon, stay tuned! 👀
• Dec 9, 2025: 🔥 Our paper on arXiv.
• Dec 2, 2025: 🎬 We released the official project page for MICo-150K.

Introduction

• We present MICo-150K, a large-scale, high-quality dataset for Multi-Image Composition (MICo) in controllable image generation. MICo focuses on synthesizing coherent and identity-consistent images from multiple reference inputs—a long-standing challenge due to the lack of suitable training data.
• MICo-150K covers 7 representative MICo tasks, constructed from carefully curated source images and diverse composition prompts. The dataset is synthesized using strong proprietary models and refined via human-in-the-loop filtering, ensuring high fidelity and identity consistency. We further introduce a Decomposition-and-Recomposition (De&Re) subset, where real-world complex images are decomposed into components and recomposed, supporting both real and synthetic compositions.
• To enable systematic evaluation, we release MICo-Bench, consisting of 1000 curated test cases, and propose Weighted-Ref-VIEScore, a new metric tailored specifically for MICo. We also provide strong baselines, including Qwen-MICo, which demonstrates competitive performance with proprietary models while supporting arbitrary multi-image inputs.

🏗️ Data Construction Pipeline

Composition Tasks (Object-Centric, Person-Centric, HOI)

We curate high-quality source images across four categories — human, object, clothes, and scene — from publicly licensed datasets, filtered and captioned by Qwen2.5-VL-72B. For each task, source images are sampled and combined using our Compose-by-Retrieval strategy: GPT-4o selects the most semantically compatible combination from candidate pools, then generates a natural composition prompt. The composite images are synthesized by Nano-Banana and verified via Qwen2.5-VL-72B (for objects/scenes) and ArcFace (for facial identity consistency).

Decompose-and-Recompose (De&Re)

We collect high-quality single-person portraits from CC12M and use Nano-Banana to decompose each into its constituent components — person, clothing, objects, and scene. Human annotators inspect and refine all decomposed components. Once verified, Nano-Banana recomposes them into a complete image. Each set of components thus yields two versions: a real-world original and a synthesized recomposition.

🔥 Qwen-MICo

Qwen-MICo is our primary baseline, fine-tuned from Qwen-Image-Edit on MICo-150K. Despite being trained on orders of magnitude less data than Qwen-Image-2509, Qwen-MICo achieves competitive or superior performance:

• Matches Qwen-Image-2509 on 3-image composition quality while supporting arbitrary numbers of input images (Qwen-Image-2509 is limited to 3).
• Produces images with higher aesthetic quality and stronger prompt adherence.
• Exhibits remarkable emergent capabilities including pose control, virtual makeup try-on, lighting transfer, and complex scene understanding — none of which were explicitly trained.

Emergent Capabilities of Qwen-MICo (click to expand)

Pose Control

Makeup Try-on

Lighting & Optics

Light Control

Phone Wallpaper

We also fine-tune four other open-source models on MICo-150K, all showing substantial improvements. Models that originally lack MICo ability (BLIP3-o, Lumina-DiMOO) acquire strong composition capabilities from scratch; models with emergent MICo ability (BAGEL, OmniGen2) are further enhanced.

📑 Open-Source Plan

• MICo-150K dataset
• MICo-Bench (897 evaluation cases + Weighted-Ref-VIEScore evaluation script)
• Finetuned Checkpoints
• Training and Inference Guidelines
• Gradio Demo
• Technical Report

🧱 Download Finetuned Models

Train

See TRAIN.md for details.

Inference

See INFER.md for details.

MICo-Bench

MICo-Bench is a comprehensive benchmark for evaluating Multi-Image Composition, containing 897 curated cases across four tasks:

We propose Weighted-Ref-VIEScore as the evaluation metric:

$$\text{Score} = W \times \text{SC} \times \text{PQ}$$

• W: Preservation score averaged over all source elements (graded ArcFace similarity for faces, binary VLM check for objects/clothes/scenes)
• SC: Semantic consistency scored by GPT-5.4 against a human-verified reference image
• PQ: Perceptual quality scored by GPT-5.4 on the generated image alone

The annotation files and evaluation script are in MICo-Bench/. The benchmark images (source images and references) are hosted on Hugging Face: 🤗 A113NW3I/MICo-Bench. See MICo-Bench/README.md for download instructions and the step-by-step evaluation guide.

🏆 Leaderboard

Please see our project page for better visualization. Feel free to raise a pull request with the bench scoring of your model 🤗

🌟 Citation

@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}
}

🙋‍♂️ Questions?

If you have any questions or suggestions, feel free to open an issue or start a discussion.