RandomBrightnessContrast

Targets:

image

volume

Image Types:uint8, float32

Adjust brightness and contrast with a torchvision-compatible default or an OpenCV-style model. Useful for training models robust to lighting variation.

Arguments

brightness_range

tuple[float, float]

[-0.2, 0.2]

Range from which the brightness adjustment is sampled. In the default torchvision-compatible mode, the multiplicative brightness factor is 1 + adjustment; both endpoints must be at least -1. In the OpenCV-style mode, the adjustment is an additive fraction of the dtype maximum. Default: (-0.2, 0.2).

contrast_range

tuple[float, float]

[-0.2, 0.2]

Range from which the contrast adjustment is sampled. The contrast factor is 1 + adjustment. In the default torchvision-compatible mode, both endpoints must be at least -1. Default: (-0.2, 0.2).

brightness_by_max

bool

false

Selects the photometric model. If False, use torchvision-compatible multiplicative brightness and mean-centered contrast. If True, use the legacy OpenCV-style affine formula with additive brightness scaled by the dtype maximum. Default: False.

ensure_safe_output

bool

false

If True, reduce the combined affine gain and offset as needed so the complete input dtype range maps inside the output range without clipping. Default: False.

p

float

0.5

Probability of applying the transform. Default: 0.5.

Examples

>>> import numpy as np
>>> import albumentations as A
>>> image = np.random.randint(0, 256, [100, 100, 3], dtype=np.uint8)

# Torchvision-compatible brightness and mean-centered contrast
>>> transform = A.RandomBrightnessContrast(p=1.0)
>>> augmented_image = transform(image=image)["image"]

# Stronger torchvision-compatible factors: brightness 0.7-1.3, contrast 0.6-1.4
>>> transform = A.RandomBrightnessContrast(
...     brightness_range=(-0.3, 0.3),
...     contrast_range=(-0.4, 0.4),
...     p=1.0,
... )
>>> augmented_image = transform(image=image)["image"]

# Legacy OpenCV-style affine model with additive brightness
>>> transform = A.RandomBrightnessContrast(
...     brightness_range=(-0.2, 0.2),
...     contrast_range=(-0.2, 0.2),
...     brightness_by_max=True,
...     p=1.0,
... )
>>> augmented_image = transform(image=image)["image"]

Notes

The default mode applies contrast first and brightness second, matching the corresponding torchvision operations in that fixed order. ColorJitter randomizes operation order.
RGB contrast uses the mean BT.601 grayscale luminance. Other multichannel inputs use the mean of their per-pixel channel averages.
A batch uses one sampled pair of factors but computes the contrast mean independently for each image. The volume target computes it independently for each slice.
Outputs are clipped to [0, 255] for uint8 and [0, 1] for float32 unless ensure_safe_output=True first reduces the affine coefficients to avoid clipping.

References

Torchvision brightness and contrast implementationhttps://docs.pytorch.org/vision/main/_modules/torchvision/transforms/v2/functional/_color.html
OpenCV basic linear transformationshttps://docs.opencv.org/4.x/d3/dc1/tutorial_basic_linear_transform.html