Migrating from 0.35 to 0.36

0.36 is a performance release with one new parameter. Nothing is removed, and every output dtype is unchanged. One result can change for an input you are unlikely to have. This page lists it; the changelog has the full list.

Results that change

equalize_histogram and perceptual_hash of a u64 or i64 image. Both work in u8, and the 64-bit integers reached it through a wrapping cast: an i64 300 became 44 and a -5 became 251. They now saturate, as every narrower integer already did (300 → 255, -5 → 0). Only values outside [0, 255] change.

New

compact=True on the numpy, torch and ndarray sinks. A tensor sink hands a row over without copying where it can, so a cropped or flipped row carries the whole storage it was cut from, and Parquet and IPC write all of it. compact=True makes each row's data exactly its elements, row-major, at offset 0, copying only rows that are not already. Use it for a column you will persist:

df.with_columns(
    crop=pl.col("image").cv.pipe(pipe).sink("ndarray", compact=True)
).write_parquet("crops.parquet")

The other sinks reject compact.

Behaviour you get for free

• Fewer copies of a view. Blur, erode, dilate, morphology gradient, pad, pad_to_size, equalize_histogram and channel_swap read a crop or a vertical flip where it lies; convolve2d and convert_color read any view without packing it first; binary ops (add, blend, apply_mask, …) read each operand block by block instead of packing both. A tensor sink that must copy a view copies it once, not twice.
• fast_image_resize 6.1, with a faster AVX2 vertical pass for u8. Resize output is unchanged.

For contributors

scripts/verify.sh now runs only what CI runs on every push (about 5 minutes on 4 cores). Add --slow for the slow lane, before a release and after engine changes. --fast is gone: it was the new default, and passing it fails with a message.