A reimplemention of the DDIM sampling / generation algorithm, as a precursor to implicit image generation via invertible DDIMs. We compare three different modes of the latest generative models, provide a library of options for noise schedulers / samplers, and analyze the best hyperparameters and model architecture decisions for generative models of various scales.
Diffusion models have emerged as a powerful class of generative models, capable of producing high-quality, diverse samples across a range of domains, including images, audio, and text. The Diffusion-IIG project specifically focuses on image generation, leveraging the invertibility and consistency of DDIM models to generate any kind of implicit image - depthmaps, segmentation maps, or albedo maps - given an arbitrary input image.
Our main goal right now is to provide a comprehensive codebase that allows easy experimentation, modification, and extension of diffusion + pixelCNN models. Stay tuned for our implicit image generation work!
To get started with Diffusion-IIG, you'll need to have Python 3.6+ and PyTorch 1.0+ installed. You can then clone this repository and follow the instructions below to setup the necessary packages.
git clone Diffusion-IIG- Run the following:
conda create Diffusion_IIG python=3.11.5 ipython
conda activate Diffusion_IIG
pip install -r requirements.txt- Implement PixelCNN benchmark. TODO: Done, integrate notebook with main repo.
- Implement DDPM benchmark.
- Noise scheduler library.
- Sampler library.
- Implement DDIM benchmark.
- UNet architecture.
- Training loop.