Landscape 64x64 image generator, based on PyTorch diffusion model.
Diffusion model that can generate random landscapes, which don`t exist on the Earth.
There is standart diffusion algorithm, which generates random noise image, and AI tries to predict which noise was used. After some linear algebra formulas we recieving some images
- Install dependencies
pip install requirements.txt
Can be problems on computer without GPU or without installed CUDA and cuDNN
After you installed dependencies, you can run generate_images.py scripts with cmd line python genarate_images.py num_of_images, where num_of_images is integer, and describes how many images will be generated (more images -> slower proccess)
Example of command: python generate_images.py 8 generates image with 8 landscapes
What will happen if we take a drawing, put some noising(not a lot), and pass it through algorithm?
This experiment probably describes an AI which redraws images in different styles.
First of all put your images in /images/example folder
You can run generate_from_images.py script with cmd line python generate_from_images.py noise_steps batch_size
noise_steps - describes how much noise will be on your image before it goes through algorithm, (I recommend to set it to 160)
batch_size - describes how many images from your dataset will be used
Example of command: python generate_from_image.py 160 4 generates image with 4 landscapes, based on 4 images in the folder
Use python app.py in cmd
It's runs on flask server on http://127.0.0.1:5000/
It's not finaly done, but somehow works( Requires page refresh to see results )