Fast-SSIM 1.3.1

Fast, multi-threaded SSIM and PSNR algorithms for Python.

Fast-SSIM

Fast, multi-threaded SSIM and PSNR calculations!

ℹ About

This fork wraps the Fast-SSIM project in an easy-to-use package that can be readily installed via PyPI.
To further accelerate processing, this version enables multithreading and leverages AVX2/FMA hardware instructions. For older CPUs it falls back to a SSE implementation.
This version achieves up to a 360x speedup for SSIM and 150x for PSNR on modern multi-core processors (tested on an AMD Ryzen 5 5600X 6-Core Processor) compared to standard implementations.

Requirements

• Python 3.9 or higher

How to Install

pip install Fast-SSIM

Usage

The functionalities are explained in the following code snippet that is also provided in this repo:

from skimage.io import imread

import fast_ssim

img1_path = r"test_images/0.jpg"
img2_path = r"test_images/1.jpg"

# Load the images into NumPy arrays
img1 = imread(img1_path)
img2 = imread(img2_path)

ssim_score = fast_ssim.ssim(
    img1, img2,
    data_range=255
)
psnr_score = fast_ssim.psnr(
    img1, img2,
    data_range=255
)

print(f"SSIM Score: {ssim_score}")
print(f"PSNR Score: {psnr_score}")

Output:

SSIM Score: 0.9886590838432312
PSNR Score: 31.11587142944336

Diagnostics & Testing

You can query the C++ backend about what hardware acceleration is being used:

import fast_ssim

status = fast_ssim.get_cpu_status()
if status == 0:
    print("Using AVX2 + FMA")
else:
    print(f"Running in compatibility mode (SSE Fallback). Status Code: {status}")

If you have a modern AVX2-capable CPU but want to test the slower SSE fallback path (for benchmarking or debugging), you can force the library to bypass AVX2 entirely.
Just set the SSIM_FORCE_SSE environment variable before importing the library:

import os
os.environ["SSIM_FORCE_SSE"] = "1"

import fast_ssim
print(fast_ssim.get_cpu_status()) # Will now output: 2 (Missing AVX2 / Forced SSE)

Performance Comparison

Due to multithreading, your exact speedup will scale dynamically with your CPU's core count. The benchmarks below were recorded on an AMD Ryzen 5 5600X 6-Core Processor:

Hardware Tested: AMD Ryzen 5 5600X 6-Core Processor

Resolution	Metric	scikit-image (Avg)	Fast-SSIM (Avg)	Speedup
1080p	SSIM	0.5804s	0.0023s	251.0x
1080p	PSNR	0.0426s	0.0002s	180.4x
4K	SSIM	2.6128s	0.0071s	367.5x
4K	PSNR	0.1782s	0.0012s	154.4x

Notes

• The SSIM calculation uses sample covariance (and sample variance) for its statistics. This aligns with the default behavior of scikit-image's structural_similarity function, where use_sample_covariance is True by default.

• This implementation does not offer an option for applying a Gaussian filter to the images or local windows prior to the SSIM/PSNR calculation.

Original author

Chen Yu / @Chen Yu