MUSAR: Exploring Multi-Subject Customization from Single-Subject Dataset via Attention Routing
Zinan Guo, Pengze Zhang, Yanze Wu✝, Chong Mou, Songtao Zhao, Qian He
(✝ Corresponding Author)
Bytedance Intelligent Creation
MUSAR-Gen Dataset
⭐️ Although MUSAR is trained solely on diptych data constructed from concatenated single-subject samples, we recognize that a high-quality multi-subject paired dataset is highly beneficial for the field of image customization. To accelerate progress in this field, we are releasing the high-quality multi-subject dataset generated by MUSAR: MUSAR-Gen. It delivers FLUX-comparable image quality without exhibiting attribute entanglement issues. Hope it will be helpful to researchers working on related topics.
Construction details: The condition images are two subjects randomly selected from the subjects200k dataset (excluding the 111,761 subjects used during the model training process). The prompt format is: "An undivided, seamless, and harmonious picture with two objects. in the xxx scene, Subject A and Subject B are placed together." By collecting the outputs of the MUSAR model, we obtained approximately 30,000 samples.
- Load dataset
from datasets import load_dataset # Load dataset dataset = load_dataset('guozinan/MUSAR-Gen')
| Key name | Type | Description |
|---|---|---|
cond_img_0 |
image |
Reference Image Information (First Image). |
cond_img_1 |
image |
Reference Image Information (Second Image). |
tgt_img |
image |
Multi-subject customized result generated by the MUSAR model. |
cond_prompt_0 |
str |
Textual description of the corresponding subject in cond_img_0. |
cond_prompt_1 |
str |
Textual description of the corresponding subject in cond_img_0. |
prompt |
str |
Textual description of the tgt_img content. |
If you find this paper useful for your research or use MUSAR-Gen dataset, please consider citing our paper:
@article{guo2025musar,
title={MUSAR: Exploring Multi-Subject Customization from Single-Subject Dataset via Attention Routing},
author={Guo, Zinan and Zhang, Pengze and Wu, Yanze and Mou, Chong and Zhao, Songtao and He, Qian},
journal={arXiv preprint arXiv:2505.02823},
year={2025}
}