This project is the result of a laboratory study exploring the primary approaches to image generation using neural networks. The developed framework provides functionality for training, comparing quality metrics, and conducting inference with two major strategies: Generative Adversarial Networks (GAN) and Denoising Diffusion Probabilistic Models (DDPM).
Alongside the main training pipelines, the project includes several modifications of vanilla methods aimed at improving final results. A detailed report of all experiments conducted based on these modifications is available in the ./docs directory.
The project is compatible with various environment managers. Examples using pipenv and conda are provided below:
pipenv
pipenv install
pipenv shell
pip install -r requirements.txt
conda
conda create -q --name image-gen-prj -c conda-forge python=3.11.10
conda activate image-gen-prj
# Install dependencies
conda install pip
pip install --upgrade pip
pip install -r requirements.txt
An automated setup script is also available:
sh setup.sh
The project provides pretrained models for inference:
- Downloading pretrained models
from huggingface_hub import snapshot_download
snapshot_download('Artem-fm/gen-ai-task', local_dir='./checkpoints')
- Running inference Examples can be found in the
./scripts/directory.
Detailed results and comparisons are documented in the following reports:
For questions or further information, please contact:
email: fedorov.am@list.ru telegram: t.me/fedorov_AMfm