Stay updated

News & Insights

Video Benchmark Results

Throughput in videos/second across video augmentation transforms. Higher is better.

CPU vs GPU — different hardware, different trade-offs

Albumentationsx runs on a single CPU thread (Apple M-series, macOS). Kornia and Torchvision run on a NVIDIA GeForce RTX 4090 with GPU-accelerated PyTorch tensors.

GPU libraries have a significant hardware advantage for throughput-heavy operations — this is expected and reflects real-world trade-offs. Use Albumentationsx when a GPU is unavailable, for transforms where it leads, or when low latency per frame matters more than raw batch throughput.

21
transforms where Albumentationsx is fastest
CPU · Apple M-series
8
transforms where Kornia is fastest
GPU · NVIDIA GeForce RTX 4090
25
transforms where Torchvision is fastest
GPU · NVIDIA GeForce RTX 4090

Speedup Distribution

How many transforms fall into each speedup range (Albumentationsx vs best competitor). 43 transforms with head-to-head comparison.

31< 0.5×20.5–1×41–2×52–5×15–10×10–50×> 50×

Results

Transform
albumentationsx
2.0.18
CPU · macOS arm64
kornia
0.8.0
NVIDIA GeForce RTX 4090
torchvision
0.21.0
NVIDIA GeForce RTX 4090
Speedup
Albx / best other
Rain25 ± 0.73.8 ± 0.06.6x
RandomGamma66 ± 2.622 ± 0.03.0x
ChannelDropout63 ± 3.422 ± 0.02.9x
Blur49 ± 2.321 ± 0.12.4x
PlankianJitter26 ± 1.111 ± 0.02.4x
MedianBlur18 ± 0.18.4 ± 0.12.2x
CornerIllumination4.9 ± 0.22.6 ± 0.11.9x
LinearIllumination5.0 ± 0.24.3 ± 0.21.2x
RGBShift25 ± 0.922 ± 0.01.1x
SaltAndPepper9.5 ± 0.28.8 ± 0.11.1x
Hue12 ± 0.920 ± 0.00.6x
CenterCrop128586 ± 7.970 ± 1.31133 ± 2350.5x
RandomCrop128529 ± 1565 ± 0.41133 ± 150.5x
GaussianIllumination6.2 ± 0.520 ± 0.10.3x
Erasing70 ± 3.7255 ± 6.60.3x
Saturation10 ± 0.837 ± 0.10.3x
ColorJitter13 ± 0.419 ± 0.069 ± 0.10.2x
GaussianNoise3.7 ± 0.122 ± 0.10.2x
Posterize69 ± 12631 ± 150.1x
Resize14 ± 1.05.9 ± 0.0140 ± 350.1x
Contrast55 ± 3.622 ± 0.0547 ± 130.1x
Invert82 ± 4.122 ± 0.2843 ± 1760.1x
Solarize60 ± 1.121 ± 0.0628 ± 5.90.1x
Grayscale73 ± 1422 ± 0.0838 ± 4670.1x
PlasmaShadow1.6 ± 0.019 ± 0.50.1x
RandomResizedCrop15 ± 0.36.3 ± 0.0182 ± 160.1x
Pad59 ± 3.1760 ± 3380.1x
PlasmaBrightness1.2 ± 0.217 ± 0.40.1x
Brightness55 ± 2.822 ± 0.0756 ± 4350.1x
VerticalFlip66 ± 2.622 ± 0.2978 ± 5.20.1x
PlasmaContrast1.1 ± 0.117 ± 0.00.1x
Sharpen26 ± 0.418 ± 0.0420 ± 9.00.1x
ChannelShuffle56 ± 2.520 ± 0.0958 ± 0.20.1x
Equalize11 ± 0.44.2 ± 0.0192 ± 1.20.1x
HorizontalFlip55 ± 1.122 ± 0.1978 ± 490.1x
GaussianBlur28 ± 0.322 ± 0.1543 ± 110.1x
Elastic5.7 ± 0.1127 ± 1.30.0x
Rotate23 ± 1.122 ± 0.1534 ± 0.20.0x
Affine16 ± 1.021 ± 0.0453 ± 0.10.0x
Perspective15 ± 0.6435 ± 0.10.0x
Normalize15 ± 0.622 ± 0.0461 ± 0.20.0x
ThinPlateSpline1.3 ± 0.145 ± 0.70.0x
AutoContrast16 ± 0.321 ± 0.0578 ± 170.0x
CLAHE8.6 ± 0.2
CoarseDropout56 ± 1.8
HSV6.2 ± 0.4
JpegCompression21 ± 0.5
LongestMaxSize18 ± 0.8
MotionBlur35 ± 1.5
OpticalDistortion11 ± 0.3
PhotoMetricDistort12 ± 0.3
Shear17 ± 0.5
SmallestMaxSize12 ± 0.1
Snow8.5 ± 0.4

Methodology

Test Environment

Albumentationsx
macOS arm64, 1 CPU thread
Kornia / Torchvision
NVIDIA GeForce RTX 4090, Linux x86_64
Videos per run
50
Runs per transform
5
Dataset
UCF101
Last run
March 11, 2025

Library Versions

Albumentationsx
2.0.18
Kornia
0.8.0
Torchvision
0.21.0
PyTorch (GPU libs)
2.6.0
Precision
torch.float16

Metric: Median throughput in videos/second across 5 runs. Higher is better.

Video loading: OpenCV for Albumentationsx; PyTorch GPU tensors (float16) for Kornia and Torchvision.

Speedup column: Albumentationsx CPU median ÷ highest GPU competitor median. Red values reflect the GPU hardware advantage, not a limitation of the library.

Want to verify the results on your own hardware or check that the comparison is fair? The benchmark code is open source on GitHub.