Learned Diffusion Processes for Lensless Cameras

For my final year project for UCL Computer Science, I investigated the use of processes from diffusion models for image reconstruction of lensless cameras, using the DiffuserCam data set.

The dataset can be found through this link: DiffuserCam Dataset

For all commands, the path to the dataset will be "path_to_dataset"

Setup

Using the environment.yml file, you can run

conda env create -f environment.yml

Or manually install the necessary using the commands below

conda create -n lensless-diffusion pytorch torchvision pytorch-cuda=11.7 -c pytorch -c nvidia
conda activate lensless-diffusion
pip install tqdm einops piq odak

Commands for training and evaluation

Models for the saved_models directory can be found through this link: Pretrained Models

By default, all models are selected if the flag is ommitted. You can select one model by adding the flag --models and its name, e.g.

--models "residual_diffusion_model_x0"

Sampling Training Images

python main.py --dataset "path_to_dataset" --sample --models "residual_diffusion_model_x0"

Sampling In The Wild Images

python main.py --dataset "path_to_dataset" --itw --models "residual_diffusion_model_x0"

Training Images

python main.py --dataset "path_to_dataset" --train --models "residual_diffusion_model_x0"

Getting average image results

Note: This will only return results if reconstructed images exist in the relevant directory

python main.py --dataset "path_to_dataset" --results --models "residual_diffusion_model_x0"