MOSAIC: Multi-Subject Personalized Generation via Correspondence-Aware Alignment and Disentanglement

Dong She^*, Siming Fu^*, Mushui Liu^*, Qiaoqiao Jin^*,
Hualiang Wang, Mu Liu, Jidong Jiang⁺
Fanqie AI Team, ByteDance

🔥 News

• 2026-01-27: 🎉 Congratulations! MOSAIC has been accepted by ICLR 2026!
• 2026-01-27: 🔥 Update the multi-subject data generation pipeline!
• 2025-09-30: 🔥 Release training/inference codes and models(resolution 512x512). The vision of resolution 1024x1024 is coming soon.
• 2025-09-30: 🔥 Release the SemAlign-MS-Subjects200K dataset
• 2025-09-02: The arXiv paper of MOSAIC is released.
• 2025-08-20: The project page of MOSAIC is released.

📖 Introduction

We present MOSAIC, a representation-centric framework that rethinks multi-subject generation through explicit semantic correspondence and orthogonal feature disentanglement. Our key insight is that multi-subject generation requires precise semantic alignment at the representation level—knowing exactly which regions in the generated image should attend to which parts of each reference.

MOSAIC introduces two key supervisions: (1) Semantic Correspondence Attention Loss (blue region) enforces precise point-to-point alignment between reference tokens and their corresponding locations in the target latent, ensuring high consistency; (2) Multi-Reference Disentanglement Loss (green region) maximizes the divergence between different references’ attention distributions, pushing each subject into orthogonal representational subspaces.

⚡️ Quick Start

🔧 Requirements and Installation

Install the requirements

conda create -n mosaic python=3.10 -y
conda activate mosaic
pip install -r requirements.txt

✍️ Inference

python inference.py 

It will download the Flux-dev model and the MOSAIC checkpoints, then execute the pre-configured demo case. You are then encouraged to experiment and create your own unique outputs by edit example_cases.json.

🚄 Training

1. Train with the SemAlign-MS-Subjects200K dataset

huggingface-cli download --repo-type dataset ByteDance-FanQie/SemAlign-MS-Subjects200K --local-dir your_data_folder/SemAlign-MS-Subjects200K

2. Train with customized dataset. To set up the preprocessing environment, first download the DIFT and GeoAware codebases. Next, integrate our preprocessing scripts by copying dift_point_matching.py and geoaware_point_matching.py into their respective directories. You can then run these scripts to generate Semantic Correspondence for your own dataset. Then you could train the code by bash train.sh

TODO

To support research and the open-source community, we will release the entire project—including datasets, inference pipelines, and model weights. Thank you for your patience and continued support! 🌟

• ✅ Release tarining and inference codes.
• ✅ Release the SemAlign-MS-Subjects200K dataset.
• ✅ Release the multi-subject driven dataset generation pipeline.
• ✅ Release model checkpoints (512x512).
• ⬜ Release model checkpoints (1024x1024).

Citation

@misc{she2025mosaicmultisubjectpersonalizedgeneration,
      title={MOSAIC: Multi-Subject Personalized Generation via Correspondence-Aware Alignment and Disentanglement}, 
      author={Dong She and Siming Fu and Mushui Liu and Qiaoqiao Jin and Hualiang Wang and Mu Liu and Jidong Jiang},
      year={2025},
      eprint={2509.01977},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.01977}, 
}