RGB Input Pipelines: Throughput, GPU Memory, and Transformation Coverage

Vladimir Iglovikov

cs.PFSep 6, 2026

RGB Input Pipelines: Throughput, GPU Memory, and Transformation Coverage

Authors: Vladimir Iglovikov

Organizations: Albumentations LLC

Abstract

An image-augmentation pipeline must deliver a complete batch before a model can use it. We compare seven input paths from five libraries, starting with RGB JPEG files and ending with a synchronized CUDA float16 batch. We manually matched transformation recipes and parameters across libraries to make the workloads as comparable as possible. The experiment uses 57 selected recipes, a batch size of 256, and one NVIDIA L4 machine. Throughput and peak process GPU memory are recorded together in 759 measurements. On the 11 recipes shared by all paths, DALI and AlbumentationsX have median throughputs of 5,029 and 4,679 images/s, with median peak GPU memory of 2,086 and 1,852 MiB. Broader pairwise comparisons favor AlbumentationsX on 26/26 TorchVision recipes, 50/51 Kornia recipes, and 25/26 Pillow recipes. DALI is faster than AlbumentationsX on all 22 shared recipes, with a median throughput ratio of 1.18x. A separate census reports coverage of the 118 entries in a selected AlbumentationsX RGB catalog. The study measures input preparation at fixed settings; it does not measure model training, numerical equivalence, or the best attainable configuration of each library. Benchmark code: https://github.com/albumentations-team/benchmark.

Figures & tables

Panel of Figure 1 — Figure 1: The input and output boundaries are shared; internal paths differ. Throughput excludes construction and warm-up. The memory window begins before construction and ends after cleanup in the same pass. Bar lengths show window placement, not measured stage durations. The initial dataset prewarm precedes this cell.

Control	Frozen setting
Input and access	10,000 sorted ImageNet-validation JPEGs; prewarmed; shared per-seed permutation
Output	CUDA float16 BCHW, $256\times 3\times 224\times 224$ ; GPU normalization
DataLoader paths	15 workers; persistent workers; pinned memory; prefetch factor 2
DALI	15 pipeline threads; native prefetch queue depth 2; native file reader and mixed decoder
Timing and repeats	One warm-up batch; 32 measured batches; seeds 137, 138, 139
Machine	Standard g2-standard-16 , one NVIDIA L4; 16 vCPUs, 64 GB host memory

Table 1: One fixed configuration is used throughout. G2 specifications identify the CPU platform as Cascade Lake [ 2 ] ; the cells record the L4 device name. Software versions and exact source identities appear in Appendix A .

Path	Decoder	Before transfer	H2D dtype	On GPU
AlbumentationsX	SimpleJPEG	uint8 recipe per image	uint8	Layout conversion and Normalize
Pillow CPU	Pillow	uint8 recipe per image; PILToTensor	uint8	Normalize
TorchVision CPU	TorchVision I/O	uint8 recipe per image	uint8	Normalize
TorchVision GPU	TorchVision I/O	Shape-forming prefix per image	uint8	Cast to float16; remaining transforms per image; stack; Normalize
Kornia CPU	TorchVision I/O	float32 recipe per image; collate; cast to float16	float16	Normalize
Kornia GPU	TorchVision I/O	float32 shape prefix per image; collate; cast to float16	float16	Remaining transforms on batch; Normalize

Table 2: Measured execution paths, verified against the source revision of the run. H2D means host-to-device transfer. Float16 conversion and normalization are separate operations; all final normalized outputs have the same contract.

Panel of Figure 2 — Figure 2: Throughput on the common set: each cell is the ratio of three-seed medians for one recipe. Blue means slower than AX; orange means faster. Colors use a symmetric log-ratio scale; printed ratios are linear. Crop means RandomCrop224; every row ends with Normalize and ToTensor. Recipe IDs link to full definitions. DALI’s Crop adds a short-side resize, and its Affine mapping differs (Section 2.3 ). Colors show effect size, not significance.

Panel of Figure 3 — Figure 3: Arithmetic mean of the per-recipe throughput ratios in Figure 2 , using the same eleven recipes for every path. Each ratio uses the median of three seeds. The dashed line marks AX at $1\times$ ; a value of $1.20\times$ means a mean relative throughput 20% higher than AX. The per-recipe values remain visible in Figure 2 .

Panel of Figure 4 — Figure 4: Median peak process GPU memory on the same eleven recipes. Each recipe contributes its median of three seeds; each bar is the median across those eleven values. Measurements exclude a model.

Panel of Figure 5 — Figure 5: AX vs Kornia: arithmetic mean of the throughput ratios on 46 recipes shared by AX, Kornia CPU, and Kornia GPU.

Panel of Figure 6 — Figure 6: AX vs TorchVision: arithmetic mean of the throughput ratios on 25 recipes shared by AX, TorchVision CPU, and TorchVision GPU.

Panel of Figure 7 — Figure 7: AX vs Pillow (PIL): arithmetic mean of the throughput ratios on 26 shared recipes. Both use their CPU paths.

Panel of Figure 8 — Figure 8: AX vs DALI: arithmetic mean of the throughput ratios on 22 shared recipes, using AX CPU and DALI GPU. The measured DALI mappings are described in Section 2.3 .

Library	Matched AX catalog entries	Census version
AlbumentationsX	118	2.3.7
Kornia	46	0.8.3
TorchVision	38	0.28.0
Pillow	23	12.3.0

Table 3: Coverage is measured relative to the selected AX catalog. These counts are neither the total API sizes of the competing libraries nor a numerical-equivalence test. DALI participates only in the execution experiment.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Implementation	Version recorded in cells
AlbumentationsX	2.4.2
Pillow CPU	12.3.0
TorchVision CPU	0.28.0+cu130
TorchVision GPU	0.28.0+cu130
Kornia CPU	0.8.3
Kornia GPU	0.8.3

Appendix

Table 4: Package versions actually recorded during execution. The dependency lock additionally fixes Torch 2.13.0+cu130, SimpleJPEG 1.9.0, and the remaining runtime packages. The CUDA-tagged Torch wheels are distinct from the base VM image.

Recipe prefix	Declared parameters and randomness scope
R01 Resize	Resize : target_size =224; interpolation ="bilinear". Sampling: deterministic.
R02 RandomCrop224	RandomCrop224 : height =224; width =224. Sampling: per-image.
R03 RandomResizedCrop	RandomResizedCrop : size =[224, 224]; width =224; scale =[0.08, 1.0]; ratio =[0.75, 1.3333333333333333]; interpolation ="bilinear". Sampling: per-image.
R04 RandomCrop224 $\to$ HorizontalFlip	RandomCrop224 : height =224; width =224. Sampling: per-image. HorizontalFlip : no parameters. Sampling: deterministic.
R05 RandomCrop224 $\to$ VerticalFlip	RandomCrop224 : height =224; width =224. Sampling: per-image. VerticalFlip : no parameters. Sampling: deterministic.
R06 Pad $\to$ RandomCrop224	Pad : padding =10; fill =0; border_mode ="constant". Sampling: deterministic. RandomCrop224 : height =224; width =224. Sampling: per-image.

Appendix

Table 13

Recipe	Throughput (images/s)			GPU memory (MiB)
	Median	Min	Max	Median	Min	Max
R01	4,751	3,576	7,759	1,852	1,852	1,852
R02	4,740	4,488	4,808	1,852	1,852	1,852
R03	4,785	4,699	4,903	1,852	1,852	1,852
R04	4,723	4,640	4,939	1,852	1,852	1,852
R05	4,907	4,859	5,060	1,852	1,852	1,852

Appendix

Table 14

Recipe	Throughput (images/s)			GPU memory (MiB)
	Median	Min	Max	Median	Min	Max
R01	2,390	1,858	2,406	1,814	1,814	1,814
R02	3,768	3,251	3,799	1,814	1,814	1,814
R03	2,721	2,662	2,788	1,814	1,814	1,814
R04	3,759	3,236	3,818	1,814	1,814	1,814
R05	3,567	3,510	3,737	1,814	1,814	1,814

Appendix

Table 15

Recipe	Throughput (images/s)			GPU memory (MiB)
	Median	Min	Max	Median	Min	Max
R01	3,671	3,577	3,736	1,814	1,814	1,814
R02	4,424	4,328	4,522	1,814	1,814	1,814
R03	3,518	3,496	3,667	1,814	1,814	1,814
R04	4,207	3,247	4,432	1,814	1,814	1,814
R05	4,319	4,094	4,363	1,814	1,814	1,814

Appendix

Table 16

Recipe	Throughput (images/s)			GPU memory (MiB)
	Median	Min	Max	Median	Min	Max
R01	3,852	3,526	3,883	1,814	1,814	1,814
R02	4,375	4,269	4,495	1,814	1,814	1,814
R03	3,558	3,509	3,604	1,814	1,814	1,814
R04	4,374	4,319	4,498	1,972	1,972	1,972
R05	4,594	4,525	4,616	1,972	1,972	1,972

Appendix

Table 17

Recipe	Throughput (images/s)			GPU memory (MiB)
	Median	Min	Max	Median	Min	Max
R01	2,063	1,961	2,636	1,776	1,776	1,776
R02	2,073	2,050	2,189	1,776	1,776	1,776
R03	1,776	1,760	1,814	1,776	1,776	1,776
R04	2,011	2,007	2,031	1,776	1,776	1,776
R05	1,980	1,531	2,007	1,776	1,776	1,776

Appendix

Table 18

Recipe	Throughput (images/s)			GPU memory (MiB)
	Median	Min	Max	Median	Min	Max
R01	1,990	1,944	2,023	1,776	1,776	1,776
R02	2,107	2,063	2,172	1,776	1,776	1,776
R03	1,746	1,446	1,799	1,776	1,776	1,776
R04	2,084	2,032	2,142	1,850	1,850	1,850
R05	2,151	2,086	2,229	1,850	1,850	1,850

Appendix

Table 19

Recipe	Throughput (images/s)			GPU memory (MiB)
	Median	Min	Max	Median	Min	Max
R01	5,070	5,010	5,076	1,766	1,766	1,958
R02	5,029	5,025	5,036	1,958	1,958	2,022
R03	5,091	5,076	5,121	2,022	2,022	2,022
R04	4,990	4,985	5,044	2,022	2,022	2,022
R05	4,992	4,969	5,041	2,022	2,022	2,022

Appendix

Table 20

AX entry	Kornia	TorchVision	Pillow
AdditiveNoise	—	—	—
AdvancedBlur	—	—	—
Affine	RandomAffine / RandomShear	RandomAffine	Image.transform(AFFINE)
AtLeastOneBBoxRandomCrop	—	RandomIoUCrop	—
AtmosphericFog	—	—	—
AutoContrast	RandomAutoContrast	RandomAutocontrast	ImageOps.autocontrast

Appendix

Table 21

Data Augmentation

Data augmentation expands or transforms training examples to improve learning, data efficiency, and robustness to distribution shifts.

See all 255 papers

AlbumentationsX: One Augmentation Pipeline for Images and Related Annotations
Aug 11, 2026Vladimir IglovikovData Augmentation
Albumentations LLC
Single-Thread JPEG Decoder Benchmarks Mis-Evaluate ML Data Loaders
May 9, 2026Vladimir Iglovikov, Dmitry Kosarevsky
Ternaus · Independent Researcher
DefaultShift: Auditing Semantic Default Shift in Accelerated Text-to-Image Models
Aug 22, 2026Xuanhua Yin, Chuanzhi Xu, Shunqi Mao +2T2I Generation Evaluation T2I Diffusion Models
School of Computer Science, The University of Sydney

Aug 11, 2026J/K move · Enter open · S save

AlbumentationsX: One Augmentation Pipeline for Images and Related Annotations

Vladimir Iglovikov

Albumentations LLC

Data Augmentation

Open paper AskPDF arXiv

Augmentation can corrupt a training example when an image and its annotations receive different random changes. A crop must use the same coordinates for the image, mask, boxes, keypoints, stereo views, video frames, or volume. Code paths that choose these values separately can silently misalign the data. AlbumentationsX keeps the transform list, probabilities, annotation settings, and random seed in one Compose object. Each call chooses random values once and applies them to every supported part of the training example. The library keeps each object's mask, box, and label together and lets projects add their own transforms. It can also save the pipeline definition, show what happened in one call, and run that call again. The examples place Compose after files have been decoded into arrays and before PyTorch groups examples into a batch. AlbumentationsX executes the declared transforms. Practitioners still decide whether a flip, crop, color change, or other operation preserves the correct label for their task.

Abstract

Data Augmentation

Explore similar work

AlbumentationsX: One Augmentation Pipeline for Images and Related Annotations

Single-Thread JPEG Decoder Benchmarks Mis-Evaluate ML Data Loaders

DefaultShift: Auditing Semantic Default Shift in Accelerated Text-to-Image Models