Installation (details)

[htoyryla]

We assume a fresh Ubuntu installation (20.04 LTS used here) with an NVIDIA driver installed as a 3-party driver during OS install.

Install miniconda from here https://docs.conda.io/en/main/miniconda.html

Then proceed:

conda create -n urdif python=3.10
conda activate urdif
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
pip install diffusers
pip install pytorch_mssim
pip install argparse
pip install einops
pip install git+https://github.com/openai/CLIP.git
pip install opencv-contrib-python
pip install kornia

Now test with (just use a path to an image in your local file system)

python urdifs.py --dir test/ --name portr2 --eta 0.5 --steps 100  --modelSize 512 --w 768  --h 768 --model unet2 --mults 1 1 2 2 4 4 8 8 --text "an abstract portrait" --textw 50 --lr 10  --load models/model-230.pt --image /home/hannu/Pictures/hft269e-adjusted.png --skip 10  --ema

Should look like this and you should get an image in the test folder.

[htoyryla]

Before trying to run urdifs.py in the example above, download the model:

Get model-230.pt to start with from https://drive.google.com/drive/folders/1s4fBeZvr23ma-a6eZMQqgeICyxiku6y_?usp=sharing and place it into folder models/

[htoyryla]

Pytorch_msssim appears to be unavailable for latest PyTorch versions. It is only needed for training using SSIM loss, so you can ignore it.

[htoyryla]

To run the GUI client (any version of urdifapp) you need to install gradio:

pip install gradio