Guided Diffusion Models

Authors: Tom LABIAUSSE - Theïlo TERRISSE

Date: Feb/Mar 2024

0 - Setup

• Clone the repository:

git clone [email protected]:t0m1ab/MVA_DELIRES_project.git

• Install delires as a package in edit mode (see config in pyproject.toml):

cd MVA_DELIRES_project/
pip install -e .

• Install python dependencies:

pip install -r requirements.txt

• Perform the data pipeline setup (nn download + kernels/masks creation + degraded datasets creation):

cd delires
bash data.sh

• Launch the benchmark:

python main.py

• Visualization of the results using the notebook visualization.ipynb

1 - Methods

1.1 - DPS [1]

1.2 - PiGDM [2]

1.3 - DiffPIR []

2 - Example of results

2.1 - Blur + Noise

2.2 - Mask + Noise

2.3 - Metrics

These metrics were obtained after benchmarking the previous methods on 100 images from the FFHQ dataset. The experimental protocol is detailed in our project report.

3 - References

• [1] Diffusion Posterior Sampling for General Noisy Inverse Problems - Chung et al. (ICLR 2023)

• [2] Pseudoinverse-Guided Diffusion Models for Inverse Problems - Song et al. (ICLR 2023)

• [3] Denoising Diffusion Models for Plug-and-Play Image Restoration - Zhu et al. (CVPR 2023)

As part of the MVA DELIRES course at ENS Paris-Saclay, this project builds on an implementation of DPS and PiGDM from Andrés ALMANSA and DiffPIR from the original authors of [1].