Nodes/ComfyUI 1hewNodes/Image HL Freq Separate
ComfyUI Node

Image HL Freq Separate

Frequency separation, the retoucher's trick, as a node

By 1hew·Created about a year ago·Updated 7 days ago· 33
Image HL Freq Separate
  • image
  • high_freq
  • low_freq
  • combine
methodrgb
blur_radius10.00

Frequency separation is the retoucher's oldest trick: split an image into a blurred "low frequency" layer (color and tone) and a "high frequency" layer (texture and fine detail), so you can smooth the skin without smearing the pores or sharpen the texture without wrecking the color. Photoshop users have done this manually for twenty years. Image HL Freq Separate is that operation, automated, in ComfyUI.

It's from ComfyUI 1hewNodes (solo dev 1hew, bilingual README, active development; the technique is inspired by the risunobushi frequency-separation pack the README credits). If you've ever wanted retouching-grade separation in a node graph, this is the cleanest entry point in the pack.

How it works

The separation is Gaussian blur-based. The low frequency output is the image blurred by blur_radius (the smooth color/tone layer). The high frequency output is the original minus the blur - remapped to sit in the middle of the 0–1 range so it shows as gray-with-detail rather than a float subtraction mess. Recombining the two should give you back the original, which is why the node also emits a combine output - a built-in sanity check that your radius is sane.

Where it gets interesting is the method dropdown:

  • rgb - the classic: blur the RGB channels. What most people mean by frequency separation.
  • hsv - blurs only the Value channel, keeping hue/saturation crisp. Better for color work where you don't want the blur dragging the hue around.
  • igbi - an inverted-gain variant that treats highlights and shadows asymmetrically. The more exotic option; produces a different feel, worth experimenting with on skin.

blur_radius is your one real knob: too small, and the "low" layer still has texture; too big, and you've flattened everything. 8–15 is a reasonable starting range for typical 1K images.

The inputs that matter

  • image - the source image (works on batches).
  • method - rgb, hsv, or igbi.
  • blur_radius - separation strength, 0–100 (default 10).

Outputs: high_freq, low_freq, and combine (the recombination check). Wire high/low into Image HL Freq Combine to rejoin with strength control, or edit either layer and recombine for retouching.

Installing it

ComfyUI Manager → search "1hew" / "ComfyUI 1hewNodes" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/1hew/ComfyUI-1hewNodes

This node needs opencv (cv2) for the Gaussian blur, which the pack installs as part of its requirements - first install is slow because the same requirements pull ultralytics/rembg/transformers for the detection nodes. No model files. Needs current ComfyUI.

Where people get burned

  • "My combine doesn't exactly match" - the recombine path applies tone-mapping levels, so it's visually equivalent, not bit-identical. Don't chase a 1:1 round trip; chase "does it look clean."
  • Halo artifacts - big blur_radius on high-contrast edges gives you halos. This is physics, not a bug; reduce the radius or use hsv.
  • Frequency separation ≠ a filter - the point is that you edit the layers between separate and combine. If you just want a blur, use a blur node. This is a pipeline, not a one-shot.
Category1hewNodes/image/hlfreq

Inputs (3)

NameTypeDefaultDescription
imageIMAGE
methodCOMBOrgb3 options: rgb, hsv, igbi
blur_radiusFLOAT10.000–100

Outputs (3)

NameTypeDescription
high_freqIMAGE
low_freqIMAGE
combineIMAGE