google-research/scenic

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

3821 stars · 480 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Scenic is a JAX-based library from Google Research focused on attention-based models for computer vision, providing shared lightweight libraries for training large-scale multi-device vision models. It includes optimized training/evaluation loops, input pipelines for popular vision datasets, and baseline implementations of SOTA models like ViViT and TokenLearner across image, video, audio, and multimodal modalities.

Use cases

• train vision transformers on multiple GPUs or TPUs
• reproduce state-of-the-art video classification models like ViViT
• build image segmentation and detection models in JAX
• run multimodal experiments combining images video and audio
• get boilerplate training loops and input pipelines for vision research
• benchmark attention-based model architectures

When to choose

• doing computer vision research with attention or transformer models
• you need multi-device multi-host training for large vision models
• you want baseline implementations of recent vision transformer papers
• your team already works in the JAX and Flax ecosystem

When to avoid

• you need a production-ready inference serving system rather than research code
• you prefer PyTorch or TensorFlow over JAX
• you want a simple off-the-shelf pretrained model API with minimal setup
• you need broad model zoo coverage outside attention-based vision research

Facets

library · maturity active

machine-learning deep-learning computer-vision image-processing video-processing audio-processing benchmarking boilerplate computer-vision deep-learning machine-learning artificial-intelligence python cross-platform jax flax vision-transformer attention transformers research-code training-loops input-pipelines multimodal google-research research video audio gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="google-research/scenic")