Kornia rebuilds its engine room in Rust

What it does

kornia-rs handles low-level computer vision grunt work—reading and writing images in over a dozen formats, resizing, color conversion, and basic tensor manipulation—all in Rust. It exposes a Python interface via PyO3 that returns numpy arrays and DLPack tensors, letting you slot the results directly into PyTorch pipelines.

The interesting bit

Instead of treating Rust as a separate ecosystem, the library uses it as a drop-in replacement for Python’s I/O and preprocessing bottlenecks. The Python API even mimics PIL’s ergonomics and supports uint16 depth maps via PNG-16, while the Rust side splits into focused sub-crates for 3D vision, AprilTags, and visual language models.

Key highlights

• Reads and writes AVIF, JPEG (via libjpeg-turbo), OpenEXR, WebP, and roughly a dozen other formats
• Thread-safe architecture, including support for Python’s free-threaded builds
• Modular crate system: kornia-io, kornia-imgproc, kornia-3d, kornia-apriltag, and others can be used independently
• Python Image class supports native uint16 for depth maps and scientific imagery, with in-memory PNG-16 encoding
• SIMD-accelerated resizing and PyTorch interoperability via DLPack

Caveats

• The Python bindings expose only a subset of the Rust API, so advanced features may require writing Rust
• Optional system dependencies like GStreamer, Video4Linux, and turbojpeg can complicate builds depending on feature flags

Verdict

Best suited for data scientists and ML engineers who need thread-safe, efficient preprocessing without leaving numpy-land. Rust developers can use the native crates directly, though the project’s center of gravity clearly leans toward Python interoperability.