An unofficial study repository for reviewing Stanford CS236 ideas and learning deep generative models, with Flow Matching as a later focus.
The basic viewpoint of this repository is to use a probabilistic model to approximate an unknown and potentially complex data distribution in the real world, such as the distribution of images, language, or other observations. Different generative models represent different approximation strategies: they choose different parameterizations, objectives, and sampling procedures for the same underlying goal of modeling data.
The notebooks are the primary learning materials. Each notebook combines a concise explanation with executable experiments, so the mathematical structure, optimization behavior, and failure modes can be inspected in one place. Toy experiments are used to make the ideas and their limitations visible before moving to higher-dimensional data.
Install the project provided in pyproject.toml. torch is sourced automatically from PyTorch's indexes: Linux x86_64 uses cu121 wheels, and other platforms fall back to CPU wheels.
Notebook experiments and supporting scripts may write files to outputs/<experiment-name>. Results are qualitative and stochastic, not benchmark claims. If imports fail, install the project in editable mode or set PYTHONPATH=src. If a run is slow, reduce the number of training steps. The experiments default to --device auto, which selects CUDA when available and falls back to CPU otherwise.