When Do Diffusion Models Learn to Generate Multiple Objects?

Yujin Jeong, Arnas Uselis, Iro Laina, Seong Joon Oh, Anna Rohrbach
ICML 2026

---

Overview

This repository contains the implementation of MOSAIC (Multi-Object Spatial relations, AttrIbution, Counting), a controlled diagnostic framework for analyzing multi-object generation failures in text-to-image diffusion models.

Key Contributions

• MOSAIC Framework: Isolates three compositional factors — Color Attribution, Counting, and Spatial Relations — enabling causal analysis of data effects.
• Comprehensive Dataset Generation: Controlled dataset generation pipeline for multi-object scenarios using Blender.
• Diagnostic Analysis: Systematic evaluation of diffusion model failures on compositional tasks.

Repository Structure

MOSAIC/
├── README.md (this file - paper overview)
├── mosaic/
│   ├── README.md (dataset generation code documentation)
│   ├── data_generation/
│   │   ├── generate_dataset.py
│   │   ├── dataset_generators.py
│   │   ├── utils.py
│   │   └── constants.py
│   ├── generation_configs/ (config examples)
│   ├── generate_dataset.sh
│   └── requirements.txt

Getting Started

For dataset generation, navigate to the mosaic folder and follow the instructions in mosaic/README.md.

TODO

• MOSAIC dataset generation code
• Training code
• Evaluation code

Citation

@inproceedings{jeong2026mosaic,
  title={When Do Diffusion Models Learn to Generate Multiple Objects?},
  author={Jeong, Yujin and Uselis, Arnas and Laina, Iro and Oh, Seong Joon and Rohrbach, Anna},
  booktitle={ICML},
  year={2026}
}