The Keras vision sidecar is merging into KerasHub

What it does KerasCV provides specialized models, layers, metrics, and callbacks for common vision tasks—classification, detection, segmentation, augmentation, and image generation—as first-party Keras objects. Built on Keras 3, components can be trained in one framework and reused in another. The library ships with pretrained weights for most models, standardized around a simple 1/255 rescaling instead of legacy ImageNet normalization schemes.

The interesting bit The README opens with a migration notice: KerasCV is being absorbed into KerasHub and will not continue as a standalone project, though existing code keeps working. That makes the repo a snapshot of a transition rather than a growing ecosystem.

Key highlights

• Native multi-backend support: write once, run on TensorFlow, JAX, or PyTorch via Keras 3.
• Pretrained weights trained entirely with KerasCV components (except StableDiffusion and the standard Vision Transformer, which carry third-party licenses).
• Standardized 1/255 image rescaling; no manual mean/std normalization required.
• Custom TensorFlow ops for 3D object detection exist but are excluded from PyPI builds to keep wheels pure-Python.

Caveats

• Active development has moved to KerasHub; this repository is effectively in maintenance mode.
• Some 3D detection features require building custom ops from source with Bazel.

Verdict Worth a look if you maintain existing KerasCV pipelines or need battle-tested, framework-agnostic CV primitives under the Keras API. Skip it if you are starting fresh—head straight to KerasHub instead.