Thanks to visit codestin.com
Credit goes to albumentations.ai

Image Benchmark Results

RGB image augmentation throughput (images/second) from the published benchmark artifacts. The DataLoader table compares every published DataLoader variant side by side; the micro benchmark isolates individual transforms.

51 of 57
DataLoader pipelines where Albumentationsx is fastest
2.4x
DataLoader average speedup
across transforms where Albumentationsx leads
7.9x
micro average speedup
vs next fastest library

Pairwise Summary

Generated automatically from published benchmark JSON at build time (median speedup midpoint with IQR).

DataLoader Pipeline

AlbumentationsX vs Kornia

Wins: 50 / 51 · Median: 3.05x · IQR: 2.76x-3.88x

AlbumentationsX vs Pillow

Wins: 26 / 26 · Median: 1.41x · IQR: 1.35x-1.70x

AlbumentationsX vs Torchvision

Wins: 26 / 26 · Median: 1.40x · IQR: 1.29x-1.83x

Micro Benchmark

AlbumentationsX vs Kornia

Wins: 51 / 51 · Median: 9.10x · IQR: 4.61x-17.06x

AlbumentationsX vs Pillow

Wins: 23 / 24 · Median: 4.56x · IQR: 3.01x-7.69x

AlbumentationsX vs Torchvision

Wins: 26 / 26 · Median: 5.61x · IQR: 3.78x-12.01x

DataLoader Pipeline Results

DataLoader benchmark with 8 workers and batch size 256. Each row is a training-style pipeline that includes crop, normalization, tensor conversion, batching, and collation. The table shows the variable transform name only. CPU/GPU memory runs are shown where published artifacts exist; DALI is a GPU run that includes JPEG decode from disk.

Speedup Distribution

How many transforms fall into each speedup range (Albumentationsx vs best competitor). 57 transforms with head-to-head comparison.

1< 0.5×50.5–1×281–2×212–5×15–10×110–50×> 50×

DataLoader Table

TransformSpeedup
Albx / best
AlbumentationsX
CPU · 2.2.6
Kornia
CPU · 0.8.2
Kornia
GPU · 0.8.2
Pillow
CPU · 12.2.0
Torchvision
CPU · 0.26.0
Torchvision
GPU · 0.26.0
DALI
GPU + decode · 2.1.0
MedianBlur12x4038 ± 089 ± 0327 ± 2167 ± 0
Elastic9.0-9.2x2878 ± 0103 ± 0316 ± 2234 ± 0118 ± 0
PlasmaBrightness4.7-5.1x2727 ± 0458 ± 0562 ± 22
PlasmaContrast4.1x2291 ± 0455 ± 0557 ± 2
MotionBlur3.9x4785 ± 01238 ± 0657 ± 2
Snow3.8x4135 ± 01102 ± 0596 ± 3
PhotoMetricDistort3.5x4269 ± 01234 ± 0580 ± 2
RandomGamma3.4x5450 ± 01624 ± 0653 ± 5
RandomRotate903.3x5087 ± 01520 ± 0632 ± 8
ChannelDropout3.0x5314 ± 01742 ± 0660 ± 3
PlasmaShadow2.9x2797 ± 0956 ± 0612 ± 22
RandomJigsaw2.9x4863 ± 01664 ± 0636 ± 2
PlankianJitter2.8x4979 ± 01763 ± 0654 ± 8
RGBShift2.8x4977 ± 01770 ± 0658 ± 3
Rain2.7x4198 ± 01541 ± 0306 ± 2
CornerIllumination2.6x3969 ± 01502 ± 0395 ± 2
GaussianIllumination2.5x3656 ± 01458 ± 0
OpticalDistortion2.5x3579 ± 01445 ± 0641 ± 6
SmallestMaxSize2.5x1370 ± 0555 ± 0179 ± 0
LongestMaxSize2.4x1363 ± 0560 ± 0179 ± 0
LinearIllumination2.4x4179 ± 01726 ± 0503 ± 2
UnsharpMask2.2x4635 ± 02129 ± 0
Sharpen2.1-2.2x4978 ± 01267 ± 0627 ± 51899 ± 02304 ± 13
EnhanceEdge1.8x4923 ± 02688 ± 0
EnhanceDetail1.7x5033 ± 02922 ± 0
Blur1.7x5507 ± 01321 ± 0651 ± 23326 ± 0
Perspective1.6x4132 ± 01325 ± 0635 ± 12640 ± 0792 ± 4
Contrast1.4-1.5x5435 ± 01829 ± 0649 ± 53244 ± 03313 ± 02282 ± 153759 ± 73
Brightness1.4x5310 ± 01832 ± 0651 ± 53551 ± 03888 ± 03097 ± 163798 ± 34
GaussianBlur1.3-1.4x5029 ± 01289 ± 0650 ± 52474 ± 02654 ± 01957 ± 143701 ± 84
Erasing1.3x5415 ± 01522 ± 0503 ± 24068 ± 02150 ± 273791 ± 45
Solarize1.3x5407 ± 01609 ± 0644 ± 23890 ± 03569 ± 04130 ± 90
Transpose1.3x5338 ± 04081 ± 0
AutoContrast1.3x4646 ± 01721 ± 0651 ± 43577 ± 02424 ± 02562 ± 78
Grayscale1.3x5378 ± 01845 ± 0655 ± 23984 ± 04188 ± 03960 ± 122
Hue1.2x4723 ± 01144 ± 0586 ± 63784 ± 5
Saturation1.2x4646 ± 01141 ± 0587 ± 43436 ± 03792 ± 30
RandomResizedCrop1.2-1.3x5056 ± 01659 ± 0905 ± 72961 ± 03870 ± 04140 ± 1823898 ± 103
Rotate1.2x4865 ± 01488 ± 0642 ± 53981 ± 03141 ± 01273 ± 83809 ± 63
ChannelShuffle1.2x5467 ± 01739 ± 0643 ± 24543 ± 03631 ± 51
Affine1.2x4528 ± 01566 ± 0646 ± 22902 ± 03175 ± 01221 ± 43806 ± 80
Equalize1.2x4511 ± 01344 ± 0321 ± 03528 ± 03062 ± 01518 ± 53824 ± 50
SaltAndPepper1.2x4508 ± 01462 ± 0375 ± 113824 ± 76
Posterize1.1-1.2x5431 ± 01660 ± 0562 ± 33897 ± 04126 ± 04726 ± 73
Invert1.1-1.2x5569 ± 01898 ± 0659 ± 33890 ± 04065 ± 04868 ± 104
ColorJiggle1.1x4255 ± 0800 ± 0529 ± 21267 ± 0626 ± 33742 ± 34
JpegCompression1.1x4267 ± 0771 ± 0590 ± 23140 ± 03447 ± 03786 ± 24
ColorJitter1.1-1.2x4299 ± 0987 ± 0573 ± 31263 ± 0629 ± 33818 ± 100
Shear1.1x4141 ± 01557 ± 02621 ± 03772 ± 54
VerticalFlip1.1x5512 ± 01887 ± 0669 ± 24078 ± 04008 ± 05020 ± 583766 ± 118
Pad1.1x4996 ± 03603 ± 04162 ± 04634 ± 363756 ± 78
HorizontalFlip1.0x5002 ± 01819 ± 0670 ± 24074 ± 04024 ± 05050 ± 583722 ± 76
CLAHE0.9-1.0x3505 ± 0789 ± 0165 ± 03731 ± 140
ThinPlateSpline0.9x743 ± 0793 ± 0598 ± 2
GaussianNoise0.9x3415 ± 01634 ± 0660 ± 73820 ± 60
RandomCrop2240.8x5130 ± 01856 ± 0959 ± 94034 ± 04367 ± 06253 ± 973589 ± 73
Resize0.4x1350 ± 0538 ± 0181 ± 0≤201235 ± 02829 ± 983523 ± 61

Micro Benchmark Results

Single-operation CPU benchmark for individual RGB transforms. This isolates augmentation cost from DataLoader batching and collation overhead.

Speedup Distribution

How many transforms fall into each speedup range (Albumentationsx vs best competitor). 57 transforms with head-to-head comparison.

< 0.5×10.5–1×81–2×202–5×155–10×1310–50×> 50×

Micro Table

TransformSpeedup
Albx / best
albumentationsx
2.2.6
kornia
0.8.2
pillow
12.2.0
torchvision
0.26.0
MedianBlur≥42x843 ± 4≤20≤20
RandomGamma32-33x9938 ± 46308 ± 2
MotionBlur24-25x1953 ± 2181 ± 1
RandomJigsaw23-24x5172 ± 16219 ± 2
Sharpen18-19x1388 ± 558 ± 075 ± 0
RandomRotate9018x5990 ± 85333 ± 4
Hue15x967 ± 166 ± 0
GaussianBlur14x2343 ± 457 ± 0169 ± 186 ± 0
ColorJiggle13-14x639 ± 534 ± 047 ± 0
PlasmaBrightness≥13x267 ± 1≤20
PhotoMetricDistort13x581 ± 445 ± 0
ColorJitter12-13x641 ± 152 ± 147 ± 0
Blur11x4449 ± 1757 ± 0409 ± 3
Elastic≥9.5x191 ± 0≤20≤20
SmallestMaxSize9.3-9.6x2017 ± 25214 ± 1
Contrast8.8-9.1x6933 ± 30771 ± 9443 ± 1475 ± 7
Brightness8.4-8.8x6912 ± 13766 ± 7609 ± 4804 ± 15
LongestMaxSize8.4-8.7x2825 ± 42330 ± 1
ChannelDropout8.1-8.4x6810 ± 65828 ± 6
PlasmaShadow7.9-8.0x420 ± 353 ± 0
Snow7.8-7.9x489 ± 362 ± 0
EnhanceDetail7.7-7.9x2148 ± 13275 ± 1
PlasmaContrast≥7.1x143 ± 0≤20
UnsharpMask6.7-6.8x906 ± 2134 ± 0
EnhanceEdge6.2-6.4x1373 ± 16219 ± 0
Invert5.4-6.1x15095 ± 611015 ± 21974 ± 262619 ± 152
Posterize5.2-5.9x14399 ± 58290 ± 91977 ± 72598 ± 137
Erasing4.9-5.3x9511 ± 74298 ± 11872 ± 71
Solarize4.9-5.0x9760 ± 34214 ± 11966 ± 6545 ± 10
SaltAndPepper4.7-4.9x738 ± 10154 ± 1
GaussianNoise4.6x225 ± 049 ± 0
RandomCrop2243.9-5.3x38380 ± 192981 ± 58492 ± 1207
CLAHE4.6x283 ± 162 ± 0
Pad4.1-4.2x13181 ± 1183167 ± 372420 ± 122
PlankianJitter3.8-3.9x2253 ± 17580 ± 2
RGBShift3.8-3.9x2292 ± 3597 ± 3
VerticalFlip3.8-3.9x14051 ± 551067 ± 23670 ± 212490 ± 149
Grayscale3.2-3.3x5194 ± 1418 ± 11591 ± 151198 ± 32
HorizontalFlip3.2-3.3x8416 ± 19920 ± 102612 ± 211999 ± 82
Perspective2.7-2.8x559 ± 2181 ± 1202 ± 2
CornerIllumination2.7x425 ± 2157 ± 0
RandomResizedCrop2.4-2.7x7150 ± 19622 ± 32823 ± 172
Resize2.4-2.6x2463 ± 37271 ± 1396 ± 4979 ± 23
Transpose2.3-2.4x4627 ± 261934 ± 35
Rain2.4x1259 ± 2527 ± 5
ChannelShuffle2.2-2.4x4337 ± 13487 ± 21866 ± 72
Affine2.1-2.2x872 ± 8402 ± 3264 ± 2240 ± 1
GaussianIllumination2.1x388 ± 1188 ± 0
Shear1.9-2.0x784 ± 6403 ± 1217 ± 0
Saturation1.7x847 ± 1767 ± 0500 ± 3
LinearIllumination1.6x521 ± 1327 ± 3
ThinPlateSpline1.4x52 ± 036 ± 0
AutoContrast1.4x1243 ± 19231 ± 1899 ± 4159 ± 0
OpticalDistortion1.4x274 ± 1201 ± 1
Rotate1.3-1.4x1408 ± 40325 ± 21045 ± 13223 ± 1
JpegCompression1.3-1.4x692 ± 743 ± 0515 ± 1512 ± 4
Equalize0.9x807 ± 3128 ± 0882 ± 12313 ± 1

Methodology

Test Environment

Platform
macOS arm64 (Apple M-series)
CPU threads
1 (forced single-thread)
Images per run
2000
Runs per transform
3
Last run
May 3, 2026

Library Versions

Albumentationsx
2.2.6
Kornia
0.8.2
Pillow
12.2.0
Torchvision
0.26.0
NumPy
2.4.4
OpenCV
4.13.0.92

Metric: Median throughput in images/second across 3 runs. Higher is better.

Warmup: Adaptive warmup until variance stabilizes before measurement begins.

Thread control: OMP, OpenBLAS, MKL, and OpenCV threads all forced to 1 to ensure fair single-thread comparison.

Image loading: Each library uses its native format — OpenCV (BGR→RGB) for Albumentationsx, normalized tensors for Kornia, PIL images for Pillow and Torchvision.

Speedup column: Range computed from throughput error bars: speedup_from to speedup_to, where each bound uses median +/- std. Green = 2x+ conservative bound, yellow = 1x-2x, gray = slower.

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