imcombiners

(Image + Combiner + Rust(rs)), made for Python.

imcombiners was built for astronomical image-stack combination. Generic reductions now live in the companion reducers package; imcombiners focuses on stack combination plus imcombine-style rejection. Pure stack reductions use reducers with imcombiners' finite-only policy: both NaN and inf are skipped.

Documentation: https://ysbach.github.io/imcombiners/ GitHub: https://github.com/ysBach/imcombiners

The package started as a tool for the main developer(@ysBach)'s reduction tools (ysfitsutilpy). It is a result of their graduate school life, TA experience (2016-2023), and astropy image combination TF experience (2020). After years of use & trial using numba, I finally rewrote the core in Rust for better speed, reliability, and maintainability.

Now it targets a modern Python API around Rust kernels, with IRAF IMCOMBINE compatibility but with better speed & API. Tests and benchmark material compare supported paths against IRAF, Astropy/NumPy, ccdproc, and bottleneck where appropriate. On a personal laptop (Apple M4 Pro), I experience a factor of few speedup over IRAF and dozens times over Python-based tools (astropy/ccdproc/bottleneck) for typical use cases. See the Documentation.

First Look

The package has four usage modes: Standard Combiner().combine() approach, compact ndcombine() wrapper, Chained Combiner(), and direct kernel calls. Start with standard Combiner usage for ordinary Python workflows. Use chained Combiner calls when you need retained diagnostics, ndcombine() for compact IRAF-like call sites (this function was made in consideration of CLI tools), and direct kernel calls for custom high-throughput layers.

import numpy as np
import imcombiners as imc

rng = np.random.default_rng(20250311)
stack = rng.normal(1000, 5, (15, 256, 256)).astype("float32")

cmb = imc.Combiner(stack)
out = cmb.combine(
    "median",  # final stack-combination method
    # 1. Optional pre-rejection threshold masking
    thresholds=(0.0, 65000.0),
    # 2. Optional per-image zero/scale normalization
    zero=None,
    scale="median",
    # 3. Optional pixel rejection before final combination
    rejectors=[
        imc.MinMaxClip(n_min=1, n_max=0.1),
        imc.SigClip(sigma=3.0, maxiters=5),
    ],
    diagnostics=None,  # output-only fast path
)

See docs/quarto/index.qmd for the detailed explanations, API-level guidance, and conventions behind this example.

Features

• Stack combination: mean, median, lower median, percentiles, sum, min, max, variance, and weighted mean via weight=.
• 1-D rejection helpers: imcombiners.kernels exposes _1d functions such as sigclip_mask_1d, pclip_1d, and minmax_combine_1d. Use the companion reducers package for standalone fast reductions such as mean, median, percentile, and variance.
• Pixel rejection: sigma, CCD noise-model, iterative linear, min/max, and IRAF-style percentile clipping. Rejection centers accept mean, median, and lower median (lmedian/lmed).
• Pipeline helpers: threshold masking, zero/scale normalization, offset padding, masks, diagnostics=None|"simple"|"full", and output-only fast paths.
• Performance docs: see docs/quarto/performance/max-performance.qmd and docs/quarto/performance/image-benchmarks.qmd.

Development Install

# You may activate your Python environment before this, e.g.,
# source ~/.venvs/your_env/bin/activate
uv pip install -e ".[dev]"

Testing and Benchmarks

uv run pytest
uv run --extra bench python benchmarks/benchmark_combine.py
uv run python benchmarks/benchmark_threads.py

--quick runs the smoke benchmark matrix. Omit it to run the full table that backs the published benchmark documentation.