A library of modular computer vision components built on Keras 3, supporting TensorFlow, JAX, and PyTorch backends.
KerasCV is a library of modular computer vision components built on Keras 3 that supports TensorFlow, JAX, and PyTorch backends. It provides production-grade, state-of-the-art workflows for tasks like data augmentation, classification, object detection, segmentation, and image generation. The library enables applied computer vision engineers to quickly assemble training and inference pipelines with components that can be trained in one framework and reused in another.
Applied computer vision engineers and machine learning practitioners who need to build production-grade vision pipelines, as well as researchers and developers working with Keras who require specialized vision components.
Developers choose KerasCV for its multi-framework support, allowing seamless switching between TensorFlow, JAX, and PyTorch, and for its modular, production-ready components that are maintained by the Keras team with strong backwards compatibility guarantees.
Industry-strength Computer Vision workflows with Keras
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Enables training and serialization in TensorFlow, JAX, or PyTorch, and reuse across frameworks without costly migrations, as highlighted in the README's multi-framework support feature.
Offers first-party Keras objects for augmentation, detection, and more with the same polish and backwards compatibility as core Keras, maintained by the Keras team.
Includes models trained using KerasCV components, with provided training scripts ensuring strong numerical performance and reproducibility, as noted in the pretrained weights section.
Standardizes on a simple 1/255 rescaling scheme, moving away from manual normalization artifacts and making pipelines more consistent, as described in the features.
The README announces a move to KerasHub, which may lead to future breaking changes or deprecation issues, creating uncertainty for long-term projects.
Advanced features like 3D object detection require custom TF ops, necessitating building from source with Bazel—a complex process compared to standard PyPI installation.
Requires setting the KERAS_BACKEND environment variable before import, which can be error-prone in multi-environment setups and adds an extra step.