Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation

Project Page | Paper

This is the official code repository for the CVPR 2024 paper titled "Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation". Please do not hesitate to reach out for any questions.

Installation Instructions

This code has been verified with python 3.9 and CUDA version 11.7. To get started, navigate to the InterpretDiffusion directory and install the necessary packages using the following commands:

git clone git@github.com:hangligit/InterpretDiffusion.git
cd InterpretDiffusion
pip install -r requirements.txt
pip install -e diffusers

Model Explanation

Our model is implemented based on the diffusers library, and we have adapted these two files specifically for our approach:

• diffusers/src/diffusers/models/unet_2d_condition.py
• diffusers/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

Demo

To see an example of how to perform inference with our pretrained concept vector, open and run the demo.ipynb notebook. We provide a set of pretrained concept vectors and the corresponding dictionary in checkpoints.

Training

The following steps illustrate how to find a concept vector, e.g., female, in a text-to-image diffusion model.

Data Generation

Run the following script to generate the training data. This takes about 1 hour for 1000 images on a single A100 GPU.

python data_creation.py

Our script allows training on a batch of concepts. Please refer to CfgBatch in data_creation.py for more details.

Training

The following script trains the concept vector. Training a single concept vector on 1000 images for 20 epochs takes approximately 2 hours on a single A100 GPU.

Please configure either wandb or tensorboard to monitor the training and validation process. Visualizing the training procedure is crucial as it indicates whether the model has learned or not.

python train.py --train_data_dir datasets/person/ --output_dir exps/exp_person

You can find a typical training procedure in the directory exps/example_person.

Testing

The following script output images for the prompt "a doctor" with the concept vector "female". For additional evaluation options, please refer to test.py

python test.py --train_data_dir datasets/person/ --output_dir exps/exp_person --num_test_samples 10 --prompt "a doctor"

To evaluate the model on the Winobias datasets, use the following command:

python test.py --train_data_dir DATASET_DIR --output_dir EXPERIMENT_DIR --evaluation_type winobias --num_test_samples 50 --template_key 0 --concept 'female' 'male' --clip_attributes 'a woman' 'a man'

For race attributes, the prompts for clip_attributes are 'a [black/white/asian]-race person'.

Citing our work

If our work has contributed to your research, we would greatly appreciate an acknowledgement by citing us as follows:

@InProceedings{li2024self,
        author    = {Li, Hang and Shen, Chengzhi and Torr, Philip and Tresp, Volker and Gu, Jindong},
        title     = {Self-discovering interpretable diffusion latent directions for responsible text-to-image generation},
        booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
        month     = {June},
        year      = {2024},
        
    }