Scaled Soft Masked Weights 🛂🅐🅒🅝

Documentation

• Class name: ScaledSoftMaskedUniversalWeights
• Category: Adv-ControlNet 🛂🅐🅒🅝/weights
• Output node: False

The ScaledSoftMaskedUniversalWeights node is designed to dynamically adjust the influence of control weights within a neural network based on a set of input parameters. It allows for the customization of weight scaling and flipping, enabling fine-tuned control over the network's behavior in response to conditional inputs or specific task requirements.

Input types

Required

• mask
  • Specifies the tensor mask that determines the spatial application of base multipliers across the control weights, enabling differentiated scaling.
  • Comfy dtype: MASK
  • Python dtype: Tensor
• min_base_multiplier
  • Sets the minimum scaling factor for the control weights, defining the lower bound of weight influence.
  • Comfy dtype: FLOAT
  • Python dtype: float
• max_base_multiplier
  • Sets the maximum scaling factor for the control weights, defining the upper bound of weight influence.
  • Comfy dtype: FLOAT
  • Python dtype: float

Optional

• uncond_multiplier
  • Specifies an unconditional scaling factor applied to the control weights, offering an additional layer of adjustment independent of other conditions.
  • Comfy dtype: FLOAT
  • Python dtype: float
• cn_extras
  • A dictionary of extra parameters for further customization of the control weights, providing a mechanism for incorporating additional, specific adjustments.
  • Comfy dtype: CN_WEIGHTS_EXTRAS
  • Python dtype: dict[str]
• autosize
  • Configures automatic sizing for the control weights, including padding adjustments, to ensure optimal fit within the network's architecture.
  • Comfy dtype: ACNAUTOSIZE
  • Python dtype: str

Output types

• CN_WEIGHTS
  • Comfy dtype: CONTROL_NET_WEIGHTS
  • The control net weights adjusted according to the input parameters, ready for use in the network.
  • Python dtype: ControlWeights
• TK_SHORTCUT
  • Comfy dtype: TIMESTEP_KEYFRAME
  • A timestep keyframe shortcut that encapsulates the adjusted control weights for temporal application within the network.
  • Python dtype: TimestepKeyframeGroup

Usage tips

• Infra type: CPU
• Common nodes: unknown

Source code

class ScaledSoftMaskedUniversalWeights:
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "mask": ("MASK", ),
                "min_base_multiplier": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.001}, ),
                "max_base_multiplier": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.001}, ),
                #"lock_min": ("BOOLEAN", {"default": False}, ),
                #"lock_max": ("BOOLEAN", {"default": False}, ),
            },
            "optional": {
                "uncond_multiplier": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}, ),
                "cn_extras": ("CN_WEIGHTS_EXTRAS",),
                "autosize": ("ACNAUTOSIZE", {"padding": 75}),
            }
        }

    RETURN_TYPES = ("CONTROL_NET_WEIGHTS", "TIMESTEP_KEYFRAME",)
    RETURN_NAMES = WEIGHTS_RETURN_NAMES
    FUNCTION = "load_weights"

    CATEGORY = "Adv-ControlNet 🛂🅐🅒🅝/weights"

    def load_weights(self, mask: Tensor, min_base_multiplier: float, max_base_multiplier: float, lock_min=False, lock_max=False,
                     uncond_multiplier: float=1.0, cn_extras: dict[str]={}):
        # normalize mask
        mask = mask.clone()
        x_min = 0.0 if lock_min else mask.min()
        x_max = 1.0 if lock_max else mask.max()
        if x_min == x_max:
            mask = torch.ones_like(mask) * max_base_multiplier
        else:
            mask = linear_conversion(mask, x_min, x_max, min_base_multiplier, max_base_multiplier)
        weights = ControlWeights.universal_mask(weight_mask=mask, uncond_multiplier=uncond_multiplier, extras=cn_extras)
        return (weights, TimestepKeyframeGroup.default(TimestepKeyframe(control_weights=weights)))