Frequency Separation

Splits an image into two layers: low (soft colour, light and shadow) and high (fine detail such as pores, hair and fabric texture). You can then fix blotchy skin tone on one layer without touching texture on the other.

In ComfyUI	Add Node → ComfyUI-JH-PixelPro → filters → GPU Frequency Separation
Good for	Skin retouching, removing colour blotches while keeping pores, texture-only sharpening or softening

When to use it

Frequency separation is the classic portrait-retouching technique:

• Low frequencies are the big, slow changes: skin colour, shading, make-up tone. This is the image blurred with a Gaussian filter.
• High frequencies are the small, fast changes: pores, stubble, eyelashes, fabric weave. This is everything the blur removed.

Retouch colour problems on low, texture problems on high, then put the two back together. A larger radius moves more of the picture into high.

Why high can be negative

high is the original minus the blurred copy. Where a pixel is brighter than its surroundings (a bright pore edge, a hair highlight) the result is positive. Where it is darker (a pore, a shadow in a crease) the result is negative. On average the values sit around zero. An image can only store values from 0 to 1, so a negative value cannot be shown as a normal picture: the preview looks dark or oddly coloured. That is expected. The pieces still add up exactly: low + high = original.

Rebuilding the image

Use Frequency Recombine to put the layers back together. Blend nodes that clamp values to 0..1 (including the layer nodes in this pack) throw the negative half of high away, so the sum no longer matches the original.

Quick start

1. Load Image → GPU Frequency Separation.
2. Leave radius = 8, sigma = 0, precision = float32.
3. Send low to Preview Image to see the smooth layer. The high preview looks dark; that is normal.
4. Edit low or high with your own nodes, then join them with Frequency Recombine. At detail = 1 it restores the original exactly.

Settings

Setting	Default	What it does	Tip
radius	8 (1 to 128)	Size of the blur in pixels. Detail smaller than this goes to high.	Small (3 to 6) for fine texture, larger for bigger blotches. Must be at most half of the shorter image side.
sigma	0.0 (0 to 50)	Blur softness. 0 means automatic: radius / 2.	Leave at 0 unless you need to match another tool.
precision	float32	float32 rebuilds the original almost exactly. float16 is about twice as fast on a modern GPU with a rebuild error around 0.001.	Use float32 for final work.

Output	Type	Meaning
low	IMAGE	The blurred, smooth layer.
high	IMAGE	The detail layer. Can contain negative values.

Tips from retouching

• Pick radius relative to image size. The same 8 px is fine detail on a 1 MP image and almost nothing on a 24 MP image.
• Smooth skin tone on low with Skin Smoother; leave high alone so pores survive.
• Use Detail Mask when you need to protect textured areas.
• Avoid float16 on CPU. It can be slow and less precise there.

Under the hood

low is a Kornia gaussian_blur2d with kernel size 2 × radius + 1, sigma radius / 2 unless overridden, and replicate borders. high = working − low in the chosen precision. Input must be 3-channel RGB in 0..1; radius above half the shorter side raises an error. A sigma larger than radius / 2 logs a warning because the kernel truncates the tail of the Gaussian. Source: core/frequency.py and nodes/frequency_node.py.

Related nodes

• Frequency Recombine
• Skin Smoother
• Smart Sharpen
• Detail Mask

Example workflow

Frequency Separation example workflow. The workflow splits the portrait, previews low and high, and rebuilds the image with Frequency Recombine at detail 1.6. It uses only core ComfyUI nodes and this pack. Copy workflows/sample_portrait.jpg into ComfyUI/input/.