Clamp Outliers

Documentation

• Class name: ClampOutliers
• Category: latent/filters
• Output node: False

The ClampOutliers node is designed to adjust the values of latent representations by clamping outliers based on a specified standard deviation threshold. This process aims to normalize the data, reducing the impact of extreme values on the overall distribution.

Input types

Required

• latents
  • The 'latents' parameter represents the latent representations to be adjusted. It is crucial for determining which data points are considered outliers and subsequently clamped.
  • Comfy dtype: LATENT
  • Python dtype: Dict[str, torch.Tensor]
• std_dev
  • The 'std_dev' parameter specifies the threshold for clamping based on standard deviations from the mean. It controls the extent to which outliers are adjusted, impacting the normalization of the data.
  • Comfy dtype: FLOAT
  • Python dtype: float

Output types

• latent
  • Comfy dtype: LATENT
  • Returns the adjusted latent representations with outliers clamped according to the specified standard deviation threshold.
  • Python dtype: Dict[str, torch.Tensor]

Usage tips

• Infra type: GPU
• Common nodes: unknown

Source code

class ClampOutliers:
    def __init__(self):
        pass

    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "latents": ("LATENT", ),
                "std_dev": ("FLOAT", {"default": 3.0, "min": 0.1, "max": 100.0, "step": 0.1,  "round": 0.1}),
            },
        }

    RETURN_TYPES = ("LATENT",)
    FUNCTION = "clamp_outliers"

    CATEGORY = "latent/filters"

    def clamp_outliers(self, latents, std_dev):
        latents_copy = copy.deepcopy(latents)
        t = latents_copy["samples"]

        for i, latent in enumerate(t):
            for j, channel in enumerate(latent):
                sd, mean = torch.std_mean(channel, dim=None)
                t[i,j] = torch.clamp(channel, min = -sd * std_dev + mean, max = sd * std_dev + mean)

        latents_copy["samples"] = t
        return (latents_copy,)