Supported Transforms

Interpolation Modes

Filter	Downscaling quality	Upscaling quality	Performance
NEAREST			⭐⭐⭐⭐⭐
BOX	⭐		⭐⭐⭐⭐
BILINEAR	⭐	⭐	⭐⭐⭐
HAMMING	⭐⭐		⭐⭐⭐
BICUBIC	⭐⭐⭐	⭐⭐⭐	⭐⭐
LANCZOS	⭐⭐⭐⭐	⭐⭐⭐⭐	⭐

Geometric Transforms

{
  "Resize": {
    "description": "Adjusts the size of the input. Supports torch.Tensor or TVTensor. Antialiasing behavior varies with input type.",
    "parameters": {
      "size": "Target output size, as an int or a sequence (h, w).",
      "interpolation": "Desired interpolation method. Possible values include BICUBIC, BILINEAR, BOX, HAMMING, LANCZOS, NEAREST. Default: BILINEAR.",
      "max_size": "(Optional) Max size for the longer edge, overruling 'size'.",
      "antialias": "(Optional) Enables antialiasing in certain modes. Default: 'None'."
    }
  },
  "ScaleJitter": {
    "description": "Applies Large Scale Jitter, suitable for various input types. Enhances data augmentation for instance segmentation.",
    "parameters": {
      "target_size": "Base scale for jittering, derived from target size dimensions.",
      "scale_range": "Scale range (min, max). Default: (0.1, 2.0).",
      "interpolation": "Desired interpolation method. Possible values include BICUBIC, BILINEAR, BOX, HAMMING, LANCZOS, NEAREST. Default: BILINEAR.",
      "antialias": "(Optional) Applies antialiasing in certain modes. Default: 'None'."
    }
  },
  "RandomShortestSize": {
    "description": "Randomly alters the shortest side of the input within set limits, maintaining aspect ratio. Suitable for tensors or TVTensors.",
    "parameters": {
      "min_size": "Minimum size for the shortest side, as an int or sequence.",
      "max_size": "(Optional) Max size for the shortest side. Default: 'None'.",
      "interpolation": "Desired interpolation method. Possible values include BICUBIC, BILINEAR, BOX, HAMMING, LANCZOS, NEAREST. Default: BILINEAR.",
      "antialias": "(Optional) Applies antialiasing in certain modes. Default: 'None'."
    }
  },
  "RandomResize": {
    "description": "Randomly resizes the input within a defined range, ideal for image segmentation augmentations.",
    "parameters": {
      "min_size": "Minimum output size for sampling.",
      "max_size": "Maximum output size for sampling.",
      "interpolation": "Desired interpolation method. Possible values include BICUBIC, BILINEAR, BOX, HAMMING, LANCZOS, NEAREST. Default: BILINEAR.",
      "antialias": "(Optional) Applies antialiasing in certain modes. Default: 'None'."
    }
  },
  "RandomCrop": {
    "description": "Crops the input randomly, supporting various input types. Useful for image segmentation.",
    "parameters": {
      "size": "Desired output size of the crop, as an int or sequence.",
      "padding": "(Optional) Padding for image borders. Default: 'None'.",
      "pad_if_needed": "Pads the image if smaller than desired size. Default: 'False'.",
      "fill": "Fill value for padding. Default: 0.",
      "padding_mode": "Padding type ('constant', 'edge', etc.). Default: 'constant'."
    }
  },
  "RandomResizedCrop": {
    "description": "Crops and resizes a random portion of the input. Commonly used in training Inception networks.",
    "parameters": {
      "size": "Output size of the crop.",
      "scale": "Bounds for random crop area. Default: (0.08, 1.0).",
      "ratio": "Bounds for aspect ratio of crop. Default: (0.75, 1.3333).",
      "interpolation": "Desired interpolation method. Possible values include BICUBIC, BILINEAR, BOX, HAMMING, LANCZOS, NEAREST. Default: BILINEAR.",
      "antialias": "(Optional) Antialiasing in certain modes. Default: 'None'."
    }
  },
  "RandomIoUCrop": {
    "description": "Random IoU-based cropping, maintaining aspect ratio and IoU with bounding boxes. Important for 'SSD: Single Shot MultiBox Detector'.",
    "parameters": {
      "min_scale": "Minimum scale factor. Default: 0.3.",
      "max_scale": "Maximum scale factor. Default: 1.0.",
      "min_aspect_ratio": "Minimum aspect ratio. Default: 0.5.",
      "max_aspect_ratio": "Maximum aspect ratio. Default: 2.0.",
      "sampler_options": "(Optional) IoU overlap values list. Default: 'None'.",
      "trials": "Trials for IoU overlap. Default: 40."
    }
  },
  "CenterCrop": {
    "description": "Crops the input from the center, suitable for various input types. Pads with 0 if input is smaller than output size.",
    "parameters": {
      "size": "Desired output size of the crop, as an int or sequence."
    }
  },
  "RandomHorizontalFlip": {
    "description": "Flips the input horizontally with a specified probability. Compatible with various input types.",
    "parameters": {
      "p": "Probability of flipping. Default: 0.5."
    }
  },
  "RandomVerticalFlip": {
    "description": "Flips the input vertically with a specified probability. Suitable for various input types.",
    "parameters": {
      "p": "Probability of flipping. Default: 0.5."
    }
  },
  "Pad": {
    "description": "Pads the input on all sides with specified padding value. Supports various input types.",
    "parameters": {
      "padding": "Padding for each border, as a single int or sequence.",
      "fill": "Fill value for padding. Default: 0.",
      "padding_mode": "Padding type ('constant', 'edge', etc.). Default: 'constant'."
    }
  },
  "RandomZoomOut": {
    "description": "Randomly zooms out on the input, creating a padding effect. Originates from 'SSD: Single Shot MultiBox Detector'.",
    "parameters": {
      "fill": "Fill value used when padding. Default: 0.",
      "side_range": "Scale factors range for zooming out. Default: [1.0, 4.0].",
      "p": "Probability of zooming out. Default: 0.5."
    }
  },
  "RandomRotation": {
    "description": "Randomly rotates the input within a specified angle range. Can expand the image to fit rotated content.",
    "parameters": {
      "degrees": "Degree range for rotation.",
      "interpolation": "Desired interpolation method. Possible values include BICUBIC, BILINEAR, BOX, HAMMING, LANCZOS, NEAREST. Default: NEAREST.",
      "expand": "Expands output to fit the rotated image. Default: 'False'.",
      "center": "(Optional) Center of rotation. Default: image center.",
      "fill": "Fill value for padding. Default: 0."
    }
  },
  "RandomAffine": {
    "description": "Applies random affine transformations (rotation, translation, scaling, shearing) while keeping the center invariant.",
    "parameters": {
      "degrees": "Degree range for rotation.",
      "translate": "(Optional) Max fraction for translations.",
      "scale": "(Optional) Scaling factor interval.",
      "shear": "(Optional) Shear degrees for x and/or y-axis.",
      "interpolation": "Desired interpolation method. Possible values include BICUBIC, BILINEAR, BOX, HAMMING, LANCZOS, NEAREST. Default: NEAREST.",
      "fill": "Fill value for padding. Default: 0.",
      "center": "(Optional) Center of rotation."
    }
  },
  "RandomPerspective": {
    "description": "Applies a random perspective transformation to the input, simulating a 3D perspective change.",
    "parameters": {
      "distortion_scale": "Degree of distortion. Default: 0.5.",
      "p": "Probability of transformation. Default: 0.5.",
      "interpolation": "Desired interpolation method. Possible values include BICUBIC, BILINEAR, BOX, HAMMING, LANCZOS, NEAREST. Default: BILINEAR.",
      "fill": "Fill value for padding. Default: 0."
    }
  },
  "ElasticTransform": {
    "description": "Generates random elastic displacements, creating a water-like effect on the input.",
    "parameters": {
      "alpha": "Magnitude of displacements. Default: 50.0.",
      "sigma": "Smoothness of displacements. Default: 5.0.",
      "interpolation": "Desired interpolation method. Possible values include BICUBIC, BILINEAR, BOX, HAMMING, LANCZOS, NEAREST. Default: BILINEAR.",
      "fill": "Fill value for padding. Default: 0."
    },
    "note": "Bounding box transformations are approximative for small displacements."
  }
}

Color Transforms

{
  "ColorJitter": {
    "description": "Randomly modifies brightness, contrast, saturation, and hue of an image or video. Not suitable for certain PIL Image modes.",
    "parameters": {
      "brightness": "(Optional) Range for brightness adjustment.",
      "contrast": "(Optional) Range for contrast adjustment.",
      "saturation": "(Optional) Range for saturation adjustment.",
      "hue": "(Optional) Single value or range for hue adjustment, between -0.5 and 0.5. If the value is a single number it must be positive."
    },
    "note": "Hue jitter requires non-negative pixel values, not suitable for normalized images with negative values."
  },
  "RandomChannelPermutation": {
    "description": "Randomly permutes the channels of an image or video. Applicable to inputs with a channel dimension."
  },
  "RandomPhotometricDistort": {
    "description": "Randomly distorts image/video's photometric properties, adjusting contrast, saturation, hue, and brightness. Includes optional channel permutation.",
    "parameters": {
      "brightness": "Range for brightness adjustment. Default: (0.875, 1.125).",
      "contrast": "Range for contrast adjustment. Default: (0.5, 1.5).",
      "saturation": "Range for saturation adjustment. Default: (0.5, 1.5).",
      "hue": "Range for hue adjustment. Default: (-0.05, 0.05).",
      "p": "Probability for each distortion. Default: 0.5."
    }
  },
  "Grayscale": {
    "description": "Converts images or videos to grayscale. Supports torch.Tensor with specific shape requirements.",
    "parameters": {
      "num_output_channels": "Desired number of output channels. Default: 1. Can be 1 or 3."
    }
  },
  "RandomGrayscale": {
    "description": "Randomly converts images or videos to grayscale with a specified probability. Maintains the number of channels in the output.",
    "parameters": {
      "p": "Probability of conversion to grayscale. Default: 0.1."
    }
  },
  "GaussianBlur": {
    "description": "Applies Gaussian blur with a randomly chosen kernel size and standard deviation. Suitable for Tensor inputs.",
    "parameters": {
      "kernel_size": "Size of the Gaussian kernel.",
      "sigma": "Standard deviation for the kernel. Default: (0.1, 2.0)."
    }
  },
  "RandomInvert": {
    "description": "Inverts the colors of an image or video randomly. Applicable to Tensor and certain PIL Image formats.",
    "parameters": {
      "p": "Probability of color inversion. Default: 0.5."
    }
  },
  "RandomPosterize": {
    "description": "Posterizes the image or video by reducing the number of bits for each color channel.",
    "parameters": {
      "bits": "Number of bits for each channel (0 to 8).",
      "p": "Probability of posterization. Default: 0.5."
    }
  },
  "RandomSolarize": {
    "description": "Solarizes the image or video by inverting pixels above a threshold.",
    "parameters": {
      "threshold": "Pixel value threshold for inversion.",
      "p": "Probability of solarization. Default: 0.5."
    }
  },
  "RandomAdjustSharpness": {
    "description": "Adjusts sharpness of an image or video randomly. Suitable for various input formats.",
    "parameters": {
      "sharpness_factor": "Level of sharpness adjustment.",
      "p": "Probability of sharpness adjustment. Default: 0.5."
    }
  },
  "RandomAutocontrast": {
    "description": "Automatically adjusts image/video contrast. Applicable to Tensor and certain PIL Image formats.",
    "parameters": {
      "p": "Probability of autocontrast. Default: 0.5."
    }
  },
  "RandomEqualize": {
    "description": "Equalizes the histogram of an image or video. Supports various input formats.",
    "parameters": {
      "p": "Probability of equalization. Default: 0.5."
    }
  }
}