Five-minute quickstart

Downsample continuous data

Use cubic-antialiasing as the default starting point for continuous-valued data that is being made smaller:

import numpy as np
from splineops import resize

image = np.random.default_rng(0).random((512, 512), dtype=np.float32)
small = resize(
    image,
    output_size=(256, 256),
    method="cubic-antialiasing",
)

Select spatial axes

When an array includes batch or channel dimensions, identify only the spatial axes. Unselected axes are preserved exactly:

# H x W x C
small_rgb = resize(
    rgb,
    output_size=(256, 256),
    axes=(0, 1),
    method="cubic-antialiasing",
)

Choose a method

Resize method selection

Goal	Method	Note
Continuous-data downsampling	cubic-antialiasing	Recommended starting point.
Faster downsampling	linear-antialiasing	Lower-order model.
Upsampling	cubic	Interpolation without projection.
Categorical labels	fast	Nearest-neighbour semantics; validate the intended grid convention.
Research configurations	resize_degrees	Exposes analysis and synthesis degrees directly.

Reuse fixed geometry

ResizePlan avoids repeating geometry setup when many arrays have the same shape:

from splineops import ResizePlan

plan = ResizePlan(
    input_shape=(64, 192, 192),
    output_size=(32, 96, 96),
    method="cubic-antialiasing",
)
coarse_frame = plan(frame)

Important semantics

SplineOps resize uses one endpoint-aligned output grid. Calls that produce the same integer output shape therefore produce the same grid and result. Review the Resize guide before comparing another library: output size, coordinates, boundaries, and antialiasing methods must all match.

Next, see Volume downsampling for a complete 3-D workflow.