pixtreme 0.3.10

pixtreme: A High-Performance Graphics Library with CUDA Support

🚀 pixtreme

Blazing-fast GPU-accelerated image processing for Python

🌟 Highlights

• ⚡ Lightning Fast: CUDA-optimized kernels deliver real-time performance
• 🎨 Professional Color Pipeline: Full ACES workflow, 3D LUTs, 10-bit precision
• 🧠 AI-Ready: Seamless integration with ONNX, PyTorch, and TensorRT
• 🔗 Zero-Copy Interop: DLPack support for PyTorch, TensorFlow, JAX
• 📊 Extensive Format Support: OpenEXR, JPEG, PNG, TIFF, and more

📋 Table of Contents

• Features
• Installation
• Quick Start
• Core Modules
• Examples
• API Reference
• Performance
• Contributing
• License

✨ Features

🎯 Image Processing

• 11 Interpolation Methods: Nearest, Linear, Cubic, Area, Lanczos (2/3/4), Mitchell, B-Spline, Catmull-Rom
• Advanced Transforms: Affine transformations, tiling with overlap blending
• Morphological Operations: Erosion with custom kernels
• GPU-Accelerated Filters: Gaussian blur, custom convolutions

🎨 Color Science

• Color Spaces: BGR/RGB, HSV, YCbCr, YUV (4:2:0, 4:2:2), Grayscale
• ACES Pipeline: Complete Academy Color Encoding System workflow
• 3D LUT Processing: Trilinear and tetrahedral interpolation
• 10-bit Precision: Professional video color accuracy

🤖 Deep Learning

• Multi-Backend Support: ONNX Runtime, PyTorch, TensorRT
• Super Resolution: Built-in upscaling with various models
• Batch Processing: Efficient multi-image inference
• Model Optimization: Automatic conversion and optimization tools

🔧 Advanced Features

• Memory I/O: Encode/decode images in memory
• Hardware Acceleration: NVIDIA nvimgcodec support
• Drawing Tools: GPU-accelerated shapes and text rendering
• Framework Integration: Zero-copy tensor sharing via DLPack

🚀 Installation

Requirements

• Python >= 3.10
• CUDA Toolkit 12.x
• NVIDIA GPU with compute capability >= 6.0

Quick Install

# Standard installation with OpenCV
pip install pixtreme[opencv]

# With OpenCV contrib modules
pip install pixtreme[opencv-contrib]

Development Setup

# Clone the repository
git clone https://github.com/sync-dev-org/pixtreme.git
cd pixtreme

# Install uv package manager
curl -LsSf https://astral.sh/uv/install.sh | sh

# Setup development environment
uv python pin 3.12
uv sync --extra dev --extra opencv

💡 Quick Start

import pixtreme as px

# Read image directly to GPU as float32 (BGR format)
image = px.imread("photo.jpg")

# All operations work on GPU memory
image_rgb = px.bgr_to_rgb(image)
image_hsv = px.rgb_to_hsv(image_rgb)

# High-quality resize with 11 interpolation methods
image = px.resize(image, (1920, 1080), interpolation=px.INTER_LANCZOS4)

# Choose backend based on your needs
upscaler = px.OnnxUpscaler("models/realesrgan.onnx")     # Balanced
# upscaler = px.TrtUpscaler("models/realesrgan.trt")     # Fastest
# upscaler = px.TorchUpscaler("models/realesrgan.pth")   # Most flexible

# Upscale with single method call
upscaled = upscaler.get(image)

# Professional color grading with 3D LUT
lut = px.read_lut("cinematic_look.cube")
graded = px.apply_lut(image, lut, interpolation=1)  # Tetrahedral

# Save with format-specific options
px.imwrite("output.jpg", graded, param=95)

📖 API Reference

🎮 Device Management

Function	Parameters	Returns	Description
get_device_id()	-	int	Get current CUDA device ID
get_device_count()	-	int	Get number of available GPUs
Device(id)	id: int	Context manager	Context manager for device selection

🎨 Color Module (pixtreme.color)

Basic Color Conversions

Function	Parameters	Returns	Description
bgr_to_rgb(image)	image: np.ndarray or cp.ndarray	np.ndarray or cp.ndarray	Convert BGR to RGB format
rgb_to_bgr(image)	image: np.ndarray or cp.ndarray	np.ndarray or cp.ndarray	Convert RGB to BGR format
bgr_to_grayscale(image)	image: cp.ndarray	cp.ndarray	Convert BGR to grayscale
rgb_to_grayscale(image)	image: cp.ndarray	cp.ndarray	Convert to grayscale (Rec.709)
bgr_to_hsv(image)	image: cp.ndarray	cp.ndarray	Convert BGR to HSV
hsv_to_bgr(image)	image: cp.ndarray	cp.ndarray	Convert HSV to BGR
rgb_to_hsv(image)	image: cp.ndarray	cp.ndarray	Convert RGB to HSV
hsv_to_rgb(image)	image: cp.ndarray	cp.ndarray	Convert HSV to RGB

YCbCr/YUV Conversions

Function	Parameters	Returns	Description
bgr_to_ycbcr(image)	image: cp.ndarray	cp.ndarray	BGR to YCbCr (10-bit precision)
rgb_to_ycbcr(image)	image: cp.ndarray	cp.ndarray	RGB to YCbCr (10-bit precision)
ycbcr_to_bgr(image)	image: cp.ndarray	cp.ndarray	YCbCr to BGR conversion
ycbcr_to_rgb(image)	image: cp.ndarray	cp.ndarray	YCbCr to RGB conversion
ycbcr_to_grayscale(image)	image: cp.ndarray	cp.ndarray	Extract Y channel as grayscale
ycbcr_full_to_legal(image)	image: cp.ndarray	cp.ndarray	Full to Legal range conversion
ycbcr_legal_to_full(image)	image: cp.ndarray	cp.ndarray	Legal to Full range conversion
yuv420p_to_ycbcr444(yuv420_data, width, height, interpolation)	yuv420_data: cp.ndarray, width: int, height: int, interpolation: int = 1	cp.ndarray	YUV 4:2:0 to YCbCr 4:4:4
yuv420p_to_ycbcr444_cp(yuv420_data, width, height, interpolation)	yuv420_data: cp.ndarray, width: int, height: int, interpolation: int = 1	cp.ndarray	YUV 4:2:0 to YCbCr 4:4:4 (CuPy native)
yuv422p10le_to_ycbcr444(ycbcr422_data, width, height)	ycbcr422_data: cp.ndarray, width: int, height: int	cp.ndarray	10-bit YUV 4:2:2 to YCbCr 4:4:4
yuv422p10le_to_ycbcr444_cp(ycbcr422_data, width, height)	ycbcr422_data: cp.ndarray, width: int, height: int	cp.ndarray	10-bit YUV 4:2:2 to YCbCr 4:4:4 (CuPy)
uyvy422_to_ycbcr444(uyvy_data, height, width)	uyvy_data: cp.ndarray, height: int, width: int	cp.ndarray	UYVY 4:2:2 to YCbCr 4:4:4
uyvy422_to_ycbcr444_cp(uyvy_data, height, width)	uyvy_data: cp.ndarray, height: int, width: int	cp.ndarray	UYVY 4:2:2 to YCbCr 4:4:4 (CuPy native)
ndi_uyvy422_to_ycbcr444(uyvy_data)	uyvy_data: cp.ndarray	cp.ndarray	NDI UYVY to YCbCr 4:4:4
ndi_uyvy422_to_ycbcr444_cp(uyvy_data)	uyvy_data: cp.ndarray	cp.ndarray	NDI UYVY to YCbCr 4:4:4 (CuPy native)

ACES Color Pipeline

Function	Parameters	Returns	Description
rec709_to_aces2065_1(image, tonemap)	image: cp.ndarray, tonemap: bool = True	cp.ndarray	Rec.709 to ACES2065-1
aces2065_1_to_rec709(image, tonemap)	image: cp.ndarray, tonemap: bool = True	cp.ndarray	ACES2065-1 to Rec.709
aces2065_1_to_acescct(image)	image: cp.ndarray	cp.ndarray	ACES2065-1 to ACEScct (log)
aces2065_1_to_acescg(image)	image: cp.ndarray	cp.ndarray	ACES2065-1 to ACEScg (linear)
acescct_to_aces2065_1(image)	image: cp.ndarray	cp.ndarray	ACEScct to ACES2065-1
acescg_to_aces2065_1(image)	image: cp.ndarray	cp.ndarray	ACEScg to ACES2065-1

3D LUT Processing

Function	Parameters	Returns	Description
read_lut(file_path, use_cache, cache_dir)	file_path: str, use_cache: bool = True, cache_dir: str = "cache"	cp.ndarray	Read .cube format LUT files
apply_lut(image, lut, interpolation)	image: cp.ndarray, lut: cp.ndarray, interpolation: int = 0	cp.ndarray	Apply 3D LUT (0=trilinear, 1=tetrahedral)
apply_lut_cp(image, lut, interpolation)	image: cp.ndarray, lut: cp.ndarray, interpolation: int = 0	cp.ndarray	Apply 3D LUT (CuPy native implementation)

🖼️ Draw Module (pixtreme.draw)

Function	Parameters	Returns	Description
circle(image, center_x, center_y, radius, color)	image: cp.ndarray, center_x: int, center_y: int, radius: int, color: tuple = (1.0, 1.0, 1.0)	cp.ndarray	Draw filled circle
rectangle(image, top_left_x, top_left_y, bottom_right_x, bottom_right_y, color)	image: cp.ndarray, top_left_x: int, top_left_y: int, bottom_right_x: int, bottom_right_y: int, color: tuple = (1.0, 1.0, 1.0)	cp.ndarray	Draw filled rectangle
put_text(image, text, org, font_face, font_scale, color, thickness, line_type, density)	image: cp.ndarray, text: str, org: tuple[int, int], font_face: int = cv2.FONT_HERSHEY_SIMPLEX, font_scale: float = 1.0, color: tuple = (1.0, 1.0, 1.0), thickness: int = 2, line_type: int = cv2.LINE_AA, density: float = 1.0	cp.ndarray	Draw text with supersampling
add_label(image, text, org, font_face, font_scale, color, thickness, line_type, label_size, label_color, label_align, density)	image: cp.ndarray, text: str, org: tuple = (0, 0), font_face: int = cv2.FONT_HERSHEY_SIMPLEX, font_scale: float = 1.0, color: tuple = (1.0, 1.0, 1.0), thickness: int = 2, line_type: int = cv2.LINE_AA, label_size: int = 20, label_color: tuple = (0.0, 0.0, 0.0), label_align: str = "bottom", density: float = 1.0	cp.ndarray	Add labeled banner
create_rounded_mask(dsize, mask_offsets, radius_ratio, density, blur_size, sigma)	dsize: tuple = (512, 512), mask_offsets: tuple = (0.1, 0.1, 0.1, 0.1), radius_ratio: float = 0.1, density: int = 1, blur_size: int = 0, sigma: float = 1.0	cp.ndarray	Create rounded rectangle mask

🔨 Filter Module (pixtreme.filter)

Function	Parameters	Returns	Description
gaussian_blur(image, kernel_size, sigma, kernel)	image: cp.ndarray, kernel_size: int, sigma: float, kernel: cp.ndarray or None = None	cp.ndarray	Apply Gaussian blur
get_gaussian_kernel(ksize, sigma)	ksize: int, sigma: float	cp.ndarray	Generate 1D Gaussian kernel
GaussianBlur(kernel_size, sigma)	Class - kernel_size: int, sigma: float	Class instance	Gaussian blur filter class

🔄 Transform Module (pixtreme.transform)

Image Operations

Function	Parameters	Returns	Description
resize(src, dsize, fx, fy, interpolation)	src: cp.ndarray, dsize: tuple[int, int] or None = None, fx: float or None = None, fy: float or None = None, interpolation: int = INTER_AUTO	cp.ndarray	Resize image with 11 interpolation methods
affine_transform(src, M, dsize, flags)	src: cp.ndarray, M: cp.ndarray, dsize: tuple, flags: int = INTER_AUTO	cp.ndarray	Apply affine transformation matrix
get_inverse_matrix(M)	M: cp.ndarray	cp.ndarray	Calculate inverse transformation matrix
crop_from_kps(image, kps, size)	image: cp.ndarray, kps: cp.ndarray, size: int = 512	tuple[cp.ndarray, cp.ndarray]	Crop image based on keypoints
erode(image, kernel_size, kernel, border_value)	image: cp.ndarray, kernel_size: int, kernel: cp.ndarray or None = None, border_value: float = 0.0	cp.ndarray	Morphological erosion
create_erode_kernel(kernel_size)	kernel_size: int	cp.ndarray	Create erosion kernel
stack_images(images, axis)	images: list[cp.ndarray], axis: int = 0	cp.ndarray	Stack multiple images
subsample_image(image, factor)	image: cp.ndarray, factor: int	cp.ndarray	Fast downsampling
subsample_image_back(image, original_shape, factor)	image: cp.ndarray, original_shape: tuple, factor: int	cp.ndarray	Upsample back to original
tile_image(input_image, tile_size, overlap)	input_image: cp.ndarray, tile_size: int = 128, overlap: int = 16	tuple[list[cp.ndarray], tuple]	Split image into tiles
merge_tiles(tiles, original_shape, padded_shape, scale, tile_size, overlap)	tiles: list[cp.ndarray], original_shape: tuple[int, int, int], padded_shape: tuple[int, int, int], scale: int, tile_size: int = 128, overlap: int = 16	cp.ndarray	Merge tiles with blending
add_padding(input_image, patch_size, overlap)	input_image: cp.ndarray, patch_size: int = 128, overlap: int = 16	cp.ndarray	Add padding for tiling
create_gaussian_weights(size, sigma)	size: int, sigma: int	cp.ndarray	Create Gaussian weight map

Interpolation Constants

• INTER_NEAREST = 0 - Nearest neighbor
• INTER_LINEAR = 1 - Bilinear
• INTER_CUBIC = 2 - Bicubic
• INTER_AREA = 3 - Area-based resampling
• INTER_LANCZOS4 = 4 - Lanczos (8x8)
• INTER_AUTO = 5 - Auto-select
• INTER_MITCHELL = 6 - Mitchell-Netravali
• INTER_B_SPLINE = 7 - B-spline
• INTER_CATMULL_ROM = 8 - Catmull-Rom
• INTER_LANCZOS2 = 9 - Lanczos (4x4)
• INTER_LANCZOS3 = 10 - Lanczos (6x6)

📁 I/O Module (pixtreme.io)

Function	Parameters	Returns	Description
imread(input_path, dtype, swap_rb, is_nvimgcodec)	input_path: str, dtype: str = "fp32", swap_rb: bool = False, is_nvimgcodec: bool = False	cp.ndarray	Read image to GPU (defaults to float32)
imwrite(output_path, image, param, swap_rb)	output_path: str, image: cp.ndarray or np.ndarray, param: int = -1, swap_rb: bool = False	None	Write image with format-specific options
imencode(image, ext, param, swap_rb)	image: cp.ndarray, ext: str = ".png", param: int = -1, swap_rb: bool = False	bytes	Encode image to memory
imdecode(src, dtype, swap_rb)	src: bytes, dtype: str = "fp32", swap_rb: bool = False	cp.ndarray	Decode image from memory
imshow(title, image, scale, is_rgb)	title: str, image: np.ndarray or cp.ndarray, scale: float = 1.0, is_rgb: bool = False	None	Display image
waitkey(delay)	delay: int	int	Wait for keyboard input
destroy_all_windows()	-	None	Close all OpenCV windows

🤖 Upscale Module (pixtreme.upscale)

Class	Constructor Parameters	Method	Description
OnnxUpscaler	model_path: str or None = None, model_bytes: bytes or None = None, device_id: int = 0, provider_options: list or None = None	get(image: cp.ndarray) -> cp.ndarray	ONNX Runtime upscaling
TrtUpscaler	model_path: str or None = None, model_bytes: bytes or None = None, device_id: int = 0	get(image: cp.ndarray) -> cp.ndarray	TensorRT optimized upscaling
TorchUpscaler	model_path: str or None = None, model_bytes: bytes or None = None, device: str = "cuda"	get(image: cp.ndarray) -> cp.ndarray	PyTorch native upscaling

🔧 Utils Module (pixtreme.utils)

Type Conversions

Function	Parameters	Returns	Description
to_uint8(image)	image: np.ndarray or cp.ndarray	np.ndarray or cp.ndarray	Convert to 8-bit (0-255)
to_uint16(image)	image: np.ndarray or cp.ndarray	np.ndarray or cp.ndarray	Convert to 16-bit unsigned
to_float16(image)	image: np.ndarray or cp.ndarray	np.ndarray or cp.ndarray	Convert to 16-bit float
to_float32(image)	image: np.ndarray or cp.ndarray	np.ndarray or cp.ndarray	Convert to 32-bit float (0-1)
to_float64(image)	image: np.ndarray or cp.ndarray	np.ndarray or cp.ndarray	Convert to 64-bit float
to_dtype(image, dtype)	image: np.ndarray or cp.ndarray, dtype: str	np.ndarray or cp.ndarray	Convert to specified dtype

Framework Interoperability

Function	Parameters	Returns	Description
to_cupy(image)	image: np.ndarray or torch.Tensor or nvimgcodec.Image	cp.ndarray	Convert to CuPy array
to_numpy(image)	image: cp.ndarray or torch.Tensor or nvimgcodec.Image	np.ndarray	Convert to NumPy array
to_tensor(image, device)	image: np.ndarray or cp.ndarray, device: str or torch.device or None = None	torch.Tensor	Convert to PyTorch tensor

Batch Processing

Function	Parameters	Returns	Description
guess_image_layout(image)	image: np.ndarray or cp.ndarray	str	Detect image layout (HW, HWC, CHW, etc.)
image_to_batch(image, size, scalefactor, mean, swap_rb, layout)	image: cp.ndarray, size: int or tuple[int, int] or None = None, scalefactor: float or None = None, mean: float or tuple or None = None, swap_rb: bool = True, layout: str = "HWC"	cp.ndarray	Convert single image to batch
images_to_batch(images, size, scalefactor, mean, swap_rb, layout)	images: list[cp.ndarray], size: int or tuple[int, int] or None = None, scalefactor: float or None = None, mean: float or tuple or None = None, swap_rb: bool = True, layout: str = "HWC"	cp.ndarray	Convert images to batch format
batch_to_images(batch, scalefactor, mean, swap_rb, layout)	batch: cp.ndarray, scalefactor: float or tuple or None = None, mean: float or tuple or None = None, swap_rb: bool = True, layout: str = "NCHW"	list[cp.ndarray]	Convert batch to images

Model Conversion

Function	Parameters	Returns	Description
check_torch_model(model_path)	model_path: str	bool	Validate PyTorch model
check_onnx_model(model_path)	model_path: str	bool	Validate ONNX model
torch_to_onnx(model_path, onnx_path, input_shape, opset_version, precision, dynamic_axes, device)	model_path: str, onnx_path: str, input_shape: tuple = (1,3,1080,1920), opset_version: int = 20, precision: str = "fp32", dynamic_axes: dict or None = None, device: str = "cuda"	None	Convert PyTorch to ONNX
onnx_to_onnx_dynamic(input_path, output_path, opset, irver)	input_path: str, output_path: str, opset: int or None = None, irver: int or None = None	None	Add dynamic shape support
onnx_to_trt(onnx_path, engine_path, precision, workspace)	onnx_path: str, engine_path: str, precision: str = "fp16", workspace: int = 1<<30	None	Convert ONNX to TensorRT
onnx_to_trt_dynamic_shape(onnx_path, engine_path, precision, workspace)	onnx_path: str, engine_path: str, precision: str = "fp16", workspace: int = 1<<30	None	TensorRT with dynamic shapes
onnx_to_trt_fixed_shape(onnx_path, engine_path, precision, workspace, input_shape)	onnx_path: str, engine_path: str, precision: str = "fp16", workspace: int = 1<<30, input_shape: tuple = (1,3,1080,1920)	None	TensorRT with fixed shape

Performance Notes

• All color conversion operations use optimized CUDA kernels
• Supports both legal range (16-235) and full range (0-255) for video processing
• 10-bit precision support for professional video workflows
• Zero-copy tensor sharing via DLPack for framework interoperability
• Batch processing support for multiple images

License

pixtreme is distributed under the MIT License (see LICENSE).

Included Components

• ACES LUTs
  • © 2014 Academy of Motion Picture Arts and Sciences
  • Licensed under "License Terms for ACES Components"
  • See third_party/aces/LICENSE_ACES.txt for details

Authors

minamik (@minamikik)

Acknowledgments

sync.dev