GitHub

Benchmarks

By avoiding global memory as much as reasonably possible we are much faster than implementations using high-level convolution calls. The Julia package SSIMLoss.jl falls under this last category, as does the Python package pytorch-msssim. On the other hand fused-ssim is a Python package (well, CUDA + other backends, with PyTorch bindings) with a similar approach to ours. Below we will compare the (CUDA) performance of these packages, as well as ImageQualityIndexes.jl as a CPU-only method for reference.

We should point out that fused_ssim.py uses zero-padding like us, while SSIMLoss.jl uses either valid padding (i.e. cropping) or symmetric padding. pytorch_msssim uses valid padding. Writing c for the number of color channels in an image, h for its width and w for its height, ImageQualityIndexes.jl and FastCUDASSIM.jl use images in (c, h, w) format, while SSIMLoss.jl uses (h, w, c), and the Python packages use (w, h, c) when converted to column-major order.

The benchmarking code can be found under benchmark in the GitHub repo. The results on a system equipped with an intel i7-7700K CPU and RTX 3070 GPU are shown below. We always present the median execution time.

  • SSIM for a single full HD image pair (1 | 3 channels):
    * ImageQualityIndexes.jl:                 242     ms  |  556     ms
    * SSIMLoss.jl (`crop = false`):            19.8   ms  |   58.2   ms
    * pytorch_msssim.py:                        4.41  ms  |    8.55  ms
    * fused_ssim.py:                            0.265 ms  |    0.870 ms
    * FastCUDASSIM.jl:                          0.271 ms  |    0.810 ms
  • DSSIMs and gradients for a batch of 32 images of size 256 x 256 (1 | 3 channels):
    * SSIMLoss.jl (Zygote):                     ERROR: Gradient Thunk(ChainRules.var"#...) should be a tuple
    * pytorch_msssim.py:                        12.6   ms  |  16.5   ms
    * fused_ssim.py:                             0.722 ms  |   1.73  ms
    * FastCUDASSIM.jl (Zygote):                  0.835 ms  |   2.17  ms
    * FastCUDASSIM.jl (`dssim_with_gradient!`):  0.629 ms  |   1.70  ms

In these benchmarks we used

  • Julia 1.12.5
  • CUDA 5.9.7 (runtime 13.1, driver 591.86.0, compiler 13.2)
  • ImageQualityIndexes 0.3.7
  • SSIMLoss 1.0.0
  • FastCUDASSIM 0.2.1
  • Python 3.10.20
  • PyTorch 2.7.1+cu118
  • pytorch_msssim 1.0.0 (latest commit: b057b07, August 16, 2023)
  • fused_ssim 1.0.0 (latest commit: a7c48d6, February 2, 2026).