EugenHotaj/pytorch-generative

Provides reference implementations of state-of-the-art generative models (autoregressive, VAEs, normalizing flows) with reusable building blocks like causal attention and layer normalization for vision transformers. Includes a unified training harness with TensorBoard integration and reproducible hyperparameters for benchmarking on standard datasets like Binarized MNIST. Supports both high-level model APIs and low-level nn components for custom architectures, with native Google Colab compatibility.

438 stars. No commits in the last 6 months.