---
Loom: >-
  Diffusion-Transformer for Interleaved Generation, and full dataset is coming
  soon.
license: apache-2.0
tags:
- >-
  multi-image reference, style transfer(sref mode), art design, painting
  process,digital art, interleaved generation.
size_categories:
- 100K<n<1M
language:
- en
pretty_name: Loom:Diffusion-Transformer for Interleaved Generation
task_categories:
- image-to-image
- text-to-image
- image-text-to-image
---
# Loom: Diffusion-Transformer for Interleaved Generation

<a href='https://arxiv.org/abs/2512.18254'><img src='https://img.shields.io/badge/Loom-Paper-green?logo=arxiv'></a>
<a href='https://github.com/Plantian/Loom'><img src='https://img.shields.io/badge/Loom-Code-blue?logo=github'></a>

![theme](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/zr4SM59gQUc9ttxv9hdQA.jpeg)

Mingcheng Ye<sup>1</sup>,
[Jiaming Liu](https://scholar.google.com/citations?hl=zh-CN&user=SmL7oMQAAAAJ&view_op=list_works&sortby=pubdate)<sup>2</sup>,
[Yiren Song](https://scholar.google.com/citations?hl=zh-CN&user=L2YS0jgAAAAJ&view_op=list_works&sortby=pubdate)<sup>3</sup><sup>\#</sup>,

\# Corresponding author

<sup>1</sup>Beijing Institute of Technology, <sup>2</sup>Alibaba Group, <sup>3</sup>National University of Singapore

## Introduction

![image](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/UDmQoJgSTQz-urz1mDURf.png)

## Dataset

### Dataset Construction

![image_7](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/OOMXO5IpkTvMfdlielHuH.png)

### Dataset Release

Full dataset is coming soon(we will release on HuggingFace).
And this is the interleaved tutorial sample belike:
![image_10](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/CuaxIjH3XOL39_4xxsbRk.png)

## More Awesome Showcases of Loom

- StyleRef Task![image_3](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/SbEW2kuzqC95Vq3N1-o7z.png)
- Multi-images Ref Task![image_4](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/Z3xS_ymr7TxvpQ-gQtaL6.png)
- Text2Interleaved Task![image_5](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/yIkfmTSyUBl76FK9kEh80.png)
- Image2Interleaved Task![image_6](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/ANld9uLSa8JEMzbPeKWPw.png)
- More![image_8](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/BfTd7MGIJNl5WRVjUaapk.png)
  ![image_9](https://cdn-uploads.huggingface.co/production/uploads/678a8b0b57a857b08f9d57c2/ynFEe1j5AufmshxYbORRd.png)

## License

Loom is licensed under the Apache 2.0.

## Citation

If you think this dataset is helpful, please reference this paper：

```bib
@misc{ye2025loomdiffusiontransformerinterleavedgeneration,
      title={Loom: Diffusion-Transformer for Interleaved Generation}, 
      author={Mingcheng Ye and Jiaming Liu and Yiren Song},
      year={2025},
      eprint={2512.18254},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.18254}, 
}
```