Noise-space HMC (N-HMC)

git clone https://github.com/Sunsett5/Noise-space-HMC.git

cd Noise-space-HMC

pip3 install gdown
gdown https://drive.google.com/uc?id=1BGwhRWUoguF-D8wlZ65tf227gp3cDUDh -O ./models/ffhq_10m.pt
gdown https://drive.google.com/uc?id=1HAy7P19PckQLczVNXmVF-e_CRxq098uW -O ./models/imagenet256.pt

Download the checkpoint "GOPRO_wVAE.pth"

gdown https://drive.google.com/uc?id=1vRoDpIsrTRYZKsOMPNbPcMtFDpCT6Foy -O ./experiments/pretrained/

Install dependencies. Change {DIR} in sed command to your root location.

conda env create -f environment.yml
conda activate NHMC
sed -i 's/torch\._six\.string_classes/str/g' /{DIR}/miniconda3/envs/NHMC/lib/python3.8/site-packages/torchvision/datasets/vision.py
sed -i "s/torch\.load(model_path, map_location='cpu')/torch\.load(model_path, map_location='cpu', weights_only=True)/" /{DIR}/miniconda3/envs/NHMC/lib/python3.8/site-packages/lpips/lpips.py

If encounter this bug "ImportError: cannot import name 'VectorQuantizer2' from 'taming.modules.vqvae.quantize'". Download quantize.py. Then replace this file miniconda/envs/NHMC/lib/python3.8/site-packages/taming/modules/vqvae/quantize.py

python3 main_sampling.py \
    --dataset ffhq \
    --timesteps 2 \
    --deg inpaint_random \
    --noise_type gaussian \
    --sigma_y 0.05 \
    --unknown_noise \
    --image_folder exp/samples/ffhq/inpaint_random \
    --verbose

• --timesteps INT
  Number of timesteps. Default: 2.

• --deg {sr4, sr16, hdr, random_inpaint, deblur_aniso, deblur_nonlinear, phase_retrieval}
  Forward operator.

• --noise_type {gaussian, speckle, impulse}
  Measurement noise type.

• --sigma_y FLOAT
  Standard deviation of measurement noise.

• --unknown_noise (flag)
  Use noise-adaptive algorithm.
  Default: False. Set to True if provided.

• --verbose (flag)
  Enable verbose output.
  Default: False.

• --image_folder PATH
  Output directory for generated images.

This repo is developed based on DPS and BlindDPS, especially for forward operations. We also use the external codes for motion-blurring and non-linear deblurring. Please also consider citing them if you use this repo.