PatchModelAddDownscale - ComfyUI Built-in Node Documentation

PatchModelAddDownscale (Kohya Deep Shrink) implements the Kohya Deep Shrink technique by applying downscaling and upscaling operations to specific blocks in a model. It reduces the resolution of intermediate features during processing and then restores them to their original size, which can improve performance while maintaining quality. The node allows precise control over when and how these scaling operations occur during the model’s execution.

Inputs

Parameter	Description	Data Type	Required	Range
`model`	The model to apply the downscale patch to	MODEL	Yes	-
`block_number`	The specific block number where downscaling will be applied (default: 3)	INT	Yes	1-32
`downscale_factor`	The factor by which to downscale the features (default: 2.0)	FLOAT	Yes	0.1-9.0
`start_percent`	The starting point in the denoising process where downscaling begins (default: 0.0)	FLOAT	Yes	0.0-1.0
`end_percent`	The ending point in the denoising process where downscaling stops (default: 0.35)	FLOAT	Yes	0.0-1.0
`downscale_after_skip`	Whether to apply downscaling after skip connections (default: True)	BOOLEAN	Yes	-
`downscale_method`	The interpolation method used for downscaling operations	COMBO	Yes	”bicubic" "nearest-exact" "bilinear" "area" "bislerp”
`upscale_method`	The interpolation method used for upscaling operations	COMBO	Yes	”bicubic" "nearest-exact" "bilinear" "area" "bislerp”

The downscale patch is applied only when the current denoising step falls within the range defined by start_percent and end_percent, and only at the block selected by block_number. When downscale_after_skip is enabled, the patch is applied after the skip connection; when disabled, it is applied before the skip connection.

Outputs

Output Name	Description	Data Type
`model`	The modified model with downscale patch applied	MODEL

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): aa9434a521ab585b290a3bd8db804469bd3bb02103a0d830b6be6eb8e8c26a5e

​Inputs

​Outputs

Inputs

Outputs