🌟 [NeurIPS 2024] Token Merging for Training-Free Semantic Binding in Text-to-Image Synthesis

📑 Introduction

Token Merging for Training-Free Semantic Binding in Text-to-Image Synthesis

Taihang Hu, Linxuan Li, Joost van de Weijer, Hongcheng Gao, Fahad Khan, Jian Yang, Ming-Ming Cheng, Kai Wang, Yaxing Wang

📚arXiv

This paper defines semantic binding as the task of associating an object with its attribute (attribute binding) or linking it to related sub-objects (object binding). We propose a novel method called Token Merging (ToMe), which enhances semantic binding by aggregating relevant tokens into a single composite token, aligning the object, its attributes, and sub-objects in the same cross-attention map.

For technical details, please refer to our paper.

🚀 Usage

1. Environment Setup

   Create and activate the Conda virtual environment:

   conda env create -f environment.yaml
   conda activate tome

   Alternatively, install dependencies via pip:

   pip install -r requirements.txt

   Additionally, download the SpaCy model for syntax parsing:

   python -m spacy download en_core_web_trf

2. Configure Parameters

   Modify the configs/demo_config.py file to adjust runtime parameters as needed. This file includes two example configuration classes: RunConfig1 for object binding and RunConfig2 for attribute binding. Key parameters are as follows:

   • prompt: Text prompt for guiding image generation.
   • model_path: Path to the Stable Diffusion model; set to None to download the pretrained model automatically.
   • use_nlp: Whether to use an NLP model for token parsing.
   • token_indices: Indices of tokens to merge.
   • prompt_anchor: Split text prompt.
   • prompt_merged: Text prompt after token merging.
   • For further parameter details, please refer to the comments in the configuration file and our paper.

3. Run the Example

   Execute the main script run_demo.py:

   python run_demo.py

   The generated images will be saved in the demo directory.

📸 Example Outputs

If everything is set up correctly, RunConfig1 and RunConfig2 should produce the left and right images below, respectively:

⚠️ Notes

• Custom Configurations: To use custom text prompts and parameters, add a new configuration class in configs/demo_config.py and make necessary adjustments in run_demo.py.
• Parameter Sensitivity: This method inherits the sensitivity of inference-based optimization techniques, meaning that the generated results are highly dependent on hyperparameter settings. Careful tuning may be required to achieve optimal results.
• NLP Models: When using NLP models like SpaCy for token parsing, ensure the correct language model is installed.

🙏 Acknowledgments

This project builds upon valuable work and resources from the following repositories:

• Attend-and-Excite
• Linguistic Binding in Diffusion Models
• 🤗 Diffusers

We extend our sincere thanks to the creators of these projects for their contributions to the field and for making their code available. 🙌