encounter1997/scenic

Scenic is a research toolkit for building and training large computer vision models. If you're working on tasks like image classification, video understanding, object detection, or combining images with audio or text, Scenic gives you pre-built components and ready-to-use example models so you don't have to start from scratch. At its core, Scenic is a collection of reusable libraries written in JAX (a numerical computing framework). It provides three main things: boilerplate code that handles the tedious infrastructure work of training models on multiple computers at scale, common building blocks like neural network layers and loss functions tailored for vision tasks, and input pipelines that efficiently load and prepare popular datasets. On top of that foundation, it includes several complete "projects", fully worked-out examples showing how to train specific models like Vision Transformers or object detectors end-to-end. Researchers and teams building computer vision systems use Scenic to move faster. Instead of reimplementing training loops, data loading, and standard model architectures, they can fork an existing baseline or project, tweak the configuration, and focus on their novel ideas. The codebase has been used to develop many published models and research papers, from video transformers to multimodal systems that combine images and audio. Scenic is designed with flexibility in mind. If you only need to adjust hyperparameters, you can change a config file and use the built-in trainers as-is. If you need deeper customization, different data pipelines, model architectures, or loss functions, you can copy and modify the relevant pieces. The README emphasizes this philosophy: the team prefers to let projects fork and adapt code rather than building overly complex abstractions, keeping everything readable and maintainable.