This is group 6's experiment on SinDDM. The modified code adds a receptive field offset argument that can be used as part of the command that runs the SinDDM main application main.py. We added a script to run the required sampling for the experiment found in here. We have also included the trained models in the repo if you desire to test the sampling yourself.
For training and other modes of sampling, please refer to the original README.
- Install Miniconda or Conda
- Create a new environment with Python 3.8
conda create --name sinddm python=3.8
conda activate sinddm
- Download requirements
pip install -r requirements.txt
Make sure you are in the sinddm conda environment before running.
python3 style_gen.py --prompt "Van Gogh Style"
If you want to sample from individual models, just replace {model}, {k}, {prompt}, {v}, and {num} to what you want to sample from.
- {model}:
cartoon,city,human_portrait,starry_night - {k}:
_small_rf,_big_rf - {prompt}: any prompt you want, for example:
Van Gogh Style - {v}:
-2(for _small_rf),0(or omit argument, for default model),2(for _big_rf) - {num}: cuda device you want to use for the sampling
python3 main.py --scope {model}{k} --mode clip_style_gen --clip_text '{prompt}' --dataset_folder ./datasets/{model}/ --image_name {model}.png --results_folder ./results/ --load_milestone 2 --receptive_field_offset {v} --device_num {num}"
Note: To run the script on a different image scale, move the desired model from the ./results/{model}_{big/small}_scale folder to the associated ./results/{model} folder, then run the model with an empty k and v of 0