---
library_name: transformers
tags:
  - vision-language-model
  - image-decomposition
---

# SynLayers

This repository contains the assets behind SynLayers, our two-stage image decomposition system.

At the root is the bbox-caption model. Given one image, it predicts:

- a whole-image caption
- bounding boxes for visible objects or layers

The same repo also includes the Stage 2 SynLayers pipeline to do layer decomposition.

If you want the easiest way to try the full system, please use our public demo:
[SynLayers/synlayers](https://huggingface.co/spaces/SynLayers/synlayers)

This repo is not meant to be used as a single generic `DiffusionPipeline(prompt)` model.
The full SynLayers pipeline is:

1. bbox + whole-caption prediction
2. layer decomposition into transparent RGBA outputs

If you only want the Stage 1 model at the repo root, you can load it with `transformers`.

```python
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "SynLayers/Bbox-caption-8b",
    torch_dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("SynLayers/Bbox-caption-8b")
```

If you want to see more details of our implementation, please check our paper out here: https://arxiv.org/abs/2605.15167

If you find our work useful, please consider citing:
@misc{wu2026doessyntheticlayereddesign,
      title={Does Synthetic Layered Design Data Benefit Layered Design Decomposition?}, 
      author={Kam Man Wu and Haolin Yang and Qingyu Chen and Yihu Tang and Jingye Chen and Qifeng Chen},
      year={2026},
      eprint={2605.15167},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.15167}, 
}

Thanks for trying SynLayers.