C.W.K.
Stream
Lesson 01 of 06 · published

What is KerasCV?

~8 min · keras-cv

Level 0Keras Apprentice
0 XP0/97 lessons0/20 achievements
0/120 XP to next level120 XP to go0% complete

KerasHub (the merged successor to KerasCV) is a model zoo with Keras-native APIs: ResNet, EfficientNet, ViT, YOLOv8, SAM, Stable Diffusion. You pull a backbone, attach a task-specific head, and you're 80% of the way to a working CV pipeline — without leaving the Keras style.

What KerasCV gives you

KerasCV is the production-grade computer vision library built on Keras 3. It bundles the pieces you'd otherwise stitch together by hand: pretrained backbones, object-detection and segmentation models with task heads already wired up, and an augmentation pipeline that runs on the same multi-backend keras.ops layer as the rest of Keras. The point isn't that any one of these is novel — it's that they all share the Keras idiom, so a model you train on JAX loads on TensorFlow for serving with no rewrite.

The KerasHub migration — read this before you import

The single most important fact about this library in 2026: KerasCV and KerasNLP were merged into one package, KerasHub. The vision models, presets, and APIs you'll meet in this track now live under keras_hub.models. The older keras_cv package still installs and runs (this track shows both import styles you'll see in the wild), but it's frozen for legacy compatibility — every new model lands in KerasHub. If you're writing fresh code, start there.

That distinction matters because most tutorials and Stack Overflow answers predate the merge. When an old keras_cv.models.X snippet errors on a model that was added after the merge, the fix is almost always "it moved to keras_hub," not "the model is gone."

External links

Exercise

Install keras-hub (pip install keras-hub). From the API docs, list 5 vision presets and note which task each is for (backbone / classifier / detector / segmenter). Load one backbone via from_preset, then print its parameter count with model.count_params(). Now reload the SAME architecture from keras_cv and confirm you get an equivalent model — proving the merge preserved behavior.
Hint
model.count_params() returns the total trainable + non-trainable parameter count. The number should match across the keras_hub and keras_cv versions of the same backbone.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.