ComfyUI_GradientDeepShrink

ComfyUI_GradientDeepShrink
★ 30

下采样线性插值Deep ShrinkPatchModel增强
基于PatchModelAddDownscale，按控制点线性插值下采样因子，方便精细调整Kohya Deep Shrink效果
💡 在训练或推理时按点平滑调整下采样强度，实现渐变缩放
🍴 5 Forks💻 Python🔄 2024-05-22
📦
网盘下载
复制链接后前往夸克网盘下载
https://pan.quark.cn/s/51137d50651f
advanced Node
📄 README

These nodes are variants of the default PatchModelAddDownscale which makes it possible to let the downscale factor scale linearly between points.

1. GradientPatchModelAddDownscale

GradientPatchModelAddDownscale modifies a given model by adding a downscaling operation to a specific block of the model. This operation is applied during the forward pass of the model.

The downscaling operation is controlled by several parameters:

  • block_number: The block in the model where the downscaling operation is applied.
  • downscale_factor: The factor by which the input is downscaled.
  • start_percent and end_percent: These parameters define a range in which the downscaling factor is linearly interpolated from 1/downscale_factor to 1.
  • downscale_after_skip: A boolean flag that determines whether the downscaling operation is applied after the skip connection in the block.

    • The node returns a new model with the downscaling operation added.

    2. GradientPatchModelAddDownscaleAdvanced

    The GradientPatchModelAddDownscaleAdvanced node is a more advanced version of the GradientPatchModelAddDownscale node. It also modifies a given model by adding a downscaling operation to a specific block of the model.

    In addition to the parameters of the GradientPatchModelAddDownscale node, the GradientPatchModelAddDownscaleAdvanced node also takes a config parameter. This parameter is a string that defines a set of percentage-scale factor pairs. The scale factor for a given percentage is determined by linearly interpolating between the defined pairs.

    The interpolate parameter is a boolean flag that determines whether the scale factor is interpolated between the defined pairs or whether the scale factor of the nearest lower percentage is used.

    The node returns a new model with the downscaling operation added.