Derivative-Free Guidance in Diffusion Models with Soft Value-Based Decoding (Images)

This code accompanies the paper on soft value-based decoding in diffusion models (SVDD), where the objective is to maximize downstream reward functions in diffusion models. In this implementation, we focus on generating images with high scores. For biological sequences, refer to here.

Nottably, our algorithm is derivative-free, training-free, and fine-tuning-free.

Code

Installation

Create a conda environment with the following command:

conda create -n SVDD_images python=3.10
conda activate SVDD_images
conda install pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

Compressibility (PM)

We use Stable Diffusion v1.5 as the pre-trained model. We optimize compressibility.

Run the following for SVDD-PM

CUDA_VISIBLE_DEVICES=0 python inference_decoding_nonp.py --reward 'compressibility' --bs 3 --num_images 3 --duplicate_size 20 --variant PM

Here is the result.

Aesthetic score (PM)

We use Stable Diffusion v1.5 as the pre-trained model. We optimize aesthetic predictors.

Run the following for SVDD-PM:

CUDA_VISIBLE_DEVICES=2 python inference_decoding_nonp.py --reward 'aesthetic' --bs 3 --num_images 3 --duplicate_size 20 --variant PM

Here is the result.

Acknowledgement

Our codebase is directly built on top of RCGDM

Reference

If you find this work useful in your research, please cite:

@article{li2024derivative,
  title={Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-Based Decoding},
  author={Li, Xiner and Zhao, Yulai and Wang, Chenyu and Scalia, Gabriele and Eraslan, Gokcen and Nair, Surag and Biancalani, Tommaso and Regev, Aviv and Levine, Sergey and Uehara, Masatoshi},
  journal={arXiv preprint arXiv:2408.08252},
  year={2024}
}