unitfield 0.2.12

High-performance N-dimensional unit field transformations with Cython-accelerated remapping

UnitField

UnitField is a high-performance N-dimensional coordinate remapping library built on a Cython kernel with OpenMP parallelism. It maps unit-space coordinates ([0, 1]) through arbitrary displacement fields with configurable interpolation and border handling — purpose-built for image warping, morphing, and nonlinear coordinate transformations.

Why UnitField?

• Cython-accelerated kernel — 2D and 1D remap loops compiled to C with OpenMP threading. Significantly faster than pure NumPy for large images.
• Asymmetric per-edge feathering — Control feather blend independently on left, right, top, and bottom borders. Useful for seamless compositing and panorama blending.
• Per-channel feather masks — Feather only specific channels (e.g., alpha-only) via the feather_dims parameter.
• Multiple border modes — CLAMP, CONSTANT, REFLECT, WRAP, REFLECT_101, and ARRAY compositing.
• Multiple interpolation methods — Nearest-neighbor, bilinear, bicubic (Catmull-Rom), Lanczos-3/4.
• 1-D signal remapping — The same kernel operates on 1-D signals, useful for audio, time-series, and look-up table applications.
• Endomorphism composition — Unit2DMappedEndomorphism supports composition (f ∘ g) for chaining transformations.
• Endomorphism interpolation — Unit1DEndoInterpolator and Unit2DEndoInterpolator blend transforms over time, even when endpoint maps use different sampled sizes.
• Fast nearest map resize — resize_unit_field_nearest upsamples or downsamples sampled remap fields for low-resolution interactive previews.

Installation

pip install unitfield

For the fastest installation with a pre-built wheel, ensure you have the cv2 extras (optional — used for comparison benchmarks only):

pip install "unitfield[cv2]"

From Source (Cython)

git clone https://github.com/Grayjou/UnitField.git
cd UnitField
pip install -e ".[dev]"

Requires: Python ≥ 3.10, NumPy ≥ 1.20, Cython ≥ 3.0 (for source builds), a C99 compiler with OpenMP support.

Quick Start

2-D Image Remapping

import numpy as np
from unitfield import (
    BorderConfig, BorderMode, InterpMethod,
    Unit2DMappedEndomorphism, remap_tensor,
)

# Simple identity field
H, W = 256, 256
xs, ys = np.meshgrid(np.linspace(0, 1, W), np.linspace(0, 1, H), indexing="xy")
identity = np.stack([xs, ys], axis=-1)
endo = Unit2DMappedEndomorphism(identity, interp_method=InterpMethod.LINEAR)

# Remap an image with asymmetric feathering
bc = BorderConfig(
    mode=BorderMode.CONSTANT,
    constant_value=0.0,
    feathering_width=0.2,
    feathering_x_overshoot_multiplier=3.0,   # heavy feather on right
    feathering_x_undershoot_multiplier=0.0,   # hard edge on left
    feather_dims=[True, True, True, False],   # RGB feathers, alpha hard
)
result = endo.remap(image, interpolation=1, border_config=bc)

Direct remap with coordinate maps

map_x = np.random.rand(H, W).astype(np.float64)
map_y = np.random.rand(H, W).astype(np.float64)

result = remap_tensor(
    image, map_x, map_y,
    interpolation=1,
    border_config=BorderConfig.constant(0.0, feathering_width=0.1),
)

Fast Preview From A Low-Resolution Map

from unitfield import InterpMethod, remap_tensor, resize_unit_field_nearest

# low_res_map shape: (preview_h, preview_w, 2), storing x/y unit coordinates
full_res_map = resize_unit_field_nearest(
    low_res_map,
    width=image.shape[1],
    height=image.shape[0],
)

preview = remap_tensor(
    image,
    full_res_map[..., 0],
    full_res_map[..., 1],
    interpolation=InterpMethod.NEAREST_MANHATTAN,
)

resize_unit_field_nearest is the preferred helper for preview remap fields. resize_nearest is also available for generic 2-D or 3-D arrays when a nearest resize is needed directly.

1-D Signal Remapping

from unitfield import remap_tensor_1d

signal = np.sin(np.linspace(0, 4 * np.pi, 1000))
map_x = np.linspace(0, 1, 800) ** 2  # nonlinear time warp
warped = remap_tensor_1d(signal, map_x, interpolation=1)

Endomorphism Interpolation

import numpy as np
from unitfield import Unit1DEndoInterpolator

ease = np.array([0.0, 0.125, 0.25, 0.5625, 1.0])
linear = np.array([0.0, 0.25, 0.5, 0.75, 1.0])

interpolator = Unit1DEndoInterpolator(ease, linear)
halfway = interpolator.interpolate(t=0.5)

value = halfway.get_value((0.5,))[0]
remapped_signal = halfway.remap(signal, interpolation=1)

Endpoint maps can be raw 1-D arrays, Unit1DMappedEndomorphism, Unit2DMappedEndomorphism, or generic UnitMappedEndomorphism instances. If endpoint maps use different sampled sizes, both are evaluated on the requested output grid before blending.

Asymmetric Feathering

BorderConfig now exposes four independent feather multipliers — one per edge:

Field	Edge	Applies when
feathering_x_undershoot_multiplier	left	u_x < 0.0
feathering_x_overshoot_multiplier	right	u_x > 1.0
feathering_y_undershoot_multiplier	top	u_y < 0.0
feathering_y_overshoot_multiplier	bottom	u_y > 1.0

All default to 1.0. Set to 0.0 for a hard edge, or higher for a softer blend.

API Overview

Module	Key exports
unitfield	BorderConfig, BorderMode, InterpMethod, remap_tensor, remap_tensor_1d, resize_unit_field_nearest, resize_nearest, Unit2DMappedEndomorphism, Unit1DMappedEndomorphism, Unit1DEndoInterpolator, Unit2DEndoInterpolator, MappedUnitField
unitfield.core	Same + UnitNdimField, UnitMappedEndomorphism, UnitEndomorphismInterpolator, UnitArray, UnitSpaceVector
unitfield.utilities	pbm_2d, upbm_2d, flat_1d_pbm — positional basematrix generators

Performance

The kernel is written in Cython with:

• OpenMP-accelerated inner loops (prange)
• No Python overhead at runtime (nogil)
• Bicubic (Catmull-Rom) and Lanczos interpolation with efficient separable sampling
• Per-edge feather distance computed inline with the border handler

Run benchmarks locally:

pytest tests/ -v -m benchmark

Development

pip install -e ".[dev]"
pytest tests/ -v

License

MIT — see LICENSE.

Citation

@software{unitfield2026,
  author = {GrayJou},
  title = {UnitField: N-dimensional Unit Field Transformations},
  year = {2026},
  url = {https://github.com/Grayjou/UnitField},
}