Scenic is a Google Research library for computer vision research, built on JAX and focused on attention-based models such as Vision Transformers. It provides the shared infrastructure and strong baselines behind many well-known vision research projects.
Scenic has two parts: a set of lightweight libraries that solve the common problems of training large vision models at scale, across multiple devices and hosts, and a collection of projects with complete training and evaluation loops for tasks like image classification, segmentation, and object detection. Although it centers on vision, its models span images, video, audio, and multimodal combinations. Being built on JAX and Flax, it fits naturally into Google’s research ecosystem and has been used to develop influential Transformer-based vision architectures and training recipes. Scenic is open source under the Apache-2.0 license, has over 3,800 stars on GitHub, and is still actively updated, with contributions continuing through 2026.
Scenic is free and open source under the Apache-2.0 license. There is no commercial version; you run it on your own compute, typically TPUs or GPUs.
Computer-vision researchers who work in the JAX ecosystem and want proven training infrastructure and baselines for attention-based models.
Scenic grew out of Google’s work on the Vision Transformer and hosts code for several influential follow-up models and training methods. Its modular design lets researchers swap backbones, losses, and data pipelines independently, which is why it is frequently chosen as the starting point for new attention-based vision experiments that need to scale from a single accelerator to large TPU pods.