InstanceAssemble

Official implementation of "InstanceAssemble: Layout-Aware Image Generation via Instance Assembling Attention" (NeurIPS 2025).

<p align="center"> <img src="fig/teaser.jpg" alt="Teaser of InstanceAssemble" width="800"> </p>

Introduction

InstanceAssemble is a lightweight framework for Layout-to-Image generation that enables precise spatial control. We also introduce DenseLayout and Layout Grounding Score (LGS) for rigorous evaluation, where InstanceAssemble achieves state-of-the-art performance on both sparse and dense layouts.

TODO

• [x] Release textual control version.
• [ ] Release additional-visual control version.
• [ ] Open-source training code.

Installation

1. Environment setup

git clone https://github.com/FireRedTeam/InstanceAssemble
cd InstanceAssemble
conda create -n instanceassemble python=3.10 -y
conda activate instanceassemble
pip install torch==2.3.0 torchvision==0.18.0 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

2. Weight Download

2.1 Download from HuggingFace

| Model Variant | Link | |---------------|:----:| | InstanceAssemble (Textual, SD3) | HuggingFace | | InstanceAssemble (Textual, Flux) | HuggingFace |

You can either download the files manually, or run:

huggingface-cli download FireRedTeam/InstanceAssemble --local-dir ./pretrained

2.2 Directory Setup

All weights should be stored under ./pretrained. A correct setup looks like:

InstanceAssemble
└── pretrained
    ├── flux
    │   ├── layout.pth
    │   └── pytorch_lora_weights.safetensors
    └── sd3
        ├── layout.pth
        └── pytorch_lora_weights.safetensors

Usage

Inference

# sd3 based
python inference.py --model_type sd3 --input_json ./demo/bigchair.json
# flux based
python inference.py --model_type fluxdev --input_json ./demo/bigchair.json
python inference.py --model_type fluxschnell --input_json ./demo/bigchair.json

Streamlit demo

streamlit run demo.py

DenseLayout and Layout Grounding Score

1. Install GroundingDINO

git clone https://github.com/IDEA-Research/GroundingDINO.git
cd GroundingDINO/
pip install -e .
mkdir weights
cd weights
wget -q https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth

For more details, please refer to GroundingDINO.

2. Generate Images by DenseLayout

python generate_dense_benchmark.py --model_type fluxdev --outdir ./output/fluxdev

3. Compute Layout Grounding Score

python score_LGS.py --imgdir ./output/fluxdev

Citation

@article{xiang2025instanceassemble,
      title={InstanceAssemble: Layout-Aware Image Generation via Instance Assembling Attention}, 
      author={Qiang Xiang and Shuang Sun and Binglei Li and Dejia Song and Huaxia Li and Nemo Chen and Xu Tang and Yao Hu and Junping Zhang},
      journal={arXiv preprint arXiv:2509.16691},
      year={2025},
}

Contact

If you have any questions about the code, please do not hesitate to contact us! Email: xiangqiang1601@163.com,sunshuang1@xiaohongshu.com