pixtreme-core

High-Performance GPU Image Processing Core Library

Overview

pixtreme-core provides the fundamental building blocks for GPU-accelerated image processing:

• I/O: Hardware-accelerated image reading/writing via NVIDIA nvimgcodec
• Color: Color space conversions (BGR, RGB, HSV, YCbCr, LUT operations)
• Transform: Geometric transformations (resize, affine, tiling) with 11 interpolation methods
• Utils: Framework interoperability (NumPy, CuPy, PyTorch) via DLPack

All operations work directly on GPU memory using CuPy arrays for maximum performance.

Installation

Requirements:

• Python >= 3.12
• CUDA Toolkit 12.x
• NVIDIA GPU with compute capability >= 6.0

pip install pixtreme-core

OpenCV Variants

pixtreme-core uses opencv-python by default. For different environments:

• Headless environments (no GUI): Replace with opencv-python-headless

  pip uninstall opencv-python
  pip install opencv-python-headless
  

• Contrib modules needed: Replace with opencv-contrib-python

  pip uninstall opencv-python
  pip install opencv-contrib-python
  

All variants provide the same cv2 module and are compatible with pixtreme.

Quick Start

import pixtreme_core as px

# Read image (returns CuPy array on GPU)
img = px.imread("input.jpg")

# Resize with auto-selected interpolation
img = px.resize(img, (512, 512))

# Convert color space
img = px.bgr_to_rgb(img)

# Write image
px.imwrite("output.jpg", img)

Features

Image I/O

• imread(): Hardware-accelerated JPEG/PNG decoding
• imwrite(): Efficient image encoding
• imshow(): Display with matplotlib

Color Conversions

• BGR ↔ RGB, HSV, YCbCr, Grayscale
• 3D LUT operations with trilinear/tetrahedral interpolation
• Video format support (UYVY422, YUV420p, YUV422p10le)
• Legal/full range YCbCr conversion

Geometric Transforms

• resize(): 11 interpolation methods including Lanczos, Mitchell, Catmull-Rom
• affine(): Affine transformations
• tile_image(), merge_tiles(): Tiling workflow for large images

Framework Interoperability

• to_cupy(), to_tensor(): GPU-to-GPU conversions via DLPack (no host round-trip)
• to_numpy(): Copy an image to a host NumPy array
• to_uint8(), to_uint16(), to_float32(): Type conversions with range scaling

License

MIT License - see LICENSE file for details.

Links

• Repository: https://github.com/sync-dev-org/pixtreme
• Documentation: https://github.com/sync-dev-org/pixtreme/blob/main/README.md