InstanceAssemble

InstanceAssemble is a lightweight layout-to-image generation framework that supports precise spatial control and enables state-of-the-art performance on sparse and dense layouts. The model was introduced at NeurIPS 2025 and introduced DenseLayout and Layout Grounding Score (LGS) for rigorous evaluation. InstanceAssemble is designed to provide flexible and efficient solutions for image generation tasks.

Demand group:

"This product is suitable for researchers and developers who require greater flexibility and control in the field of image generation. InstanceAssemble 's efficiency and accuracy make it an important tool for a new generation of image generation tasks."

Example of usage scenario:

In interior design, generate high-quality interior scenes based on a given layout.

Use this model to generate product display diagrams for e-commerce platforms.

Provide rapid scene generation support for animation production.

Product features:

Supports image generation for both sparse and dense layouts.

Introduce Layout Grounding Score for evaluation.

Model variants based on text and visual controls are available.

Supports fast inference and image generation.

Compatible with HuggingFace platform, easy to download and use models.

Usage tutorial:

Clone the project locally: git clone https://github.com/FireRedTeam/InstanceAssemble

Create and activate the environment: conda create -n InstanceAssemble python=3.10 -y && conda activate InstanceAssemble

Install dependencies: pip install -r requirements.txt

Download the model weights to the pretrained directory: huggingface-cli download FireRedTeam/ InstanceAssemble --local-dir ./pretrained

Run the inference script for image generation: python inference.py --model_type sd3 --input_json ./demo/bigchair.json