A tile texture generater using Denoising Diffusion Probabilistic Models (DDPM) with Classifier-free Guidance and DDIM sampling.
pip install -r requirements.txt
If you have trouble installing PyYAML, try
pip install --ignore-installed PyYAML
The project uses following folder structure for dataset.
dataset/
└── images/
├── 00/
│ ├── 00.jpg
│ ├── ...
├── 01/
│ ├── ...
├── ...
Make sure that the configuration files in config/ meet your needs, or create a new one instead.
View config/template.yaml for more information.
config/
├── template.yaml
├── config_64.yaml
├── config_128.yaml
└── [your new configuration file here]
Then run
python train.py --config [your config]
If you want to resume training, add --load_latest
For example, to resume training a 64*64 model
python train.py --config config/config_64.yaml --load_latest
python sample.py --model_path [your model path]
You can optionally add the following command parameters
--steps [steps]: sampling steps. Less is faster--eta [eta]: eta, 0.0 for deterministic sampler--batches [batches]: The number of batches to sample--w [w]: guided w mention in Classifier-free Diffusion Guidance--mode [DDPM/DDIM]: To choose which sampler to use
To control the categories of generated images, change object y in sample.py yourself.
For instance, set y to [0, 2, 5, 30] means generate four pictures per batch, and the categories are 0, 2, 5 and 30.
A demo built with Flask. Install Flask first.
pip install Flask
Then run
python app.py
and access 127.0.0.1/Image_Generate.
We have released models here used for demo.
The codes are heavily based on https://github.com/FutureXiang/Diffusion. The model parameters are set with reference to Beat GANs.