RC High/Low Frequency Skin Smoothing

The reason so much AI "portrait retouching" looks like a mannequin is that it smooths everything uniformly - pores, edges, eyelashes, all of it - until the face has the texture of a ping-pong ball. Frequency separation exists to avoid exactly that. You split the image into a smooth low-frequency layer (color and tone) and a detail high-frequency layer (pores, texture), smooth only the low frequencies, and put the detail back on top. Skin gets softer; edges and texture stay.

RC High/Low Frequency Skin Smoothing automates that whole pipeline with skin-tone detection and edge protection built in. It's the rare one-click retouching node that understands the difference between a blemish and an eyebrow.

How it works

The core method (frequency_separation) is textbook: blur the image with a Gaussian at low_frequency_radius to get the low-frequency layer, subtract it from the original to isolate high-frequency detail, smooth the low-frequency layer, and recombine. The high_frequency_strength slider controls how much of that detail comes back - 0 is maximum smoothing, 1.0 restores original detail, above 1.0 actually sharpens texture.

Around that core sit the safeguards:

• skin_tone_protection - masks the effect toward detected skin tones. At 0.8 (default), 80% of the smoothing budget is reserved for skin, leaving hair, clothing, and background alone.
• preserve_fine_details - keeps eyelashes, eyebrows, and hair from being smoothed into mush.
• edge_protection - preserves hard edges (jawlines, lips) from blur bleed.
• method - you can also pick surface_blur or selective_gaussian, two alternative smoothing approaches that some find more forgiving on texture. Frequency separation is the classic for a reason, though.
• mask (optional) - a MASK input to limit smoothing to specific regions. White = smooth, black = preserve. If the auto skin detection misses, this is your override.

Output is a single IMAGE tensor.

How to install it

Part of the RC Image Compositor pack:

cd ComfyUI/custom_nodes
git clone https://github.com/kj863257/ComfyUI_RC_Image_Compositor
cd ComfyUI_RC_Image_Compositor
pip install -r requirements.txt

Restart ComfyUI, or install via ComfyUI Manager by searching "RC Image Compositor". Deps: pillow, numpy, opencv-python.

Common issues

The biggest practical mistake is cranking high_frequency_strength to 0 and expecting beauty - at 0 you've thrown away all texture, and even with edge protection the result is waxy. Keep it at 0.5–0.7 for a "still human" retouch and only drop lower for a deliberate airbrush look.

Second, low_frequency_radius controls what counts as "texture" vs. "tone." Too small (1–3) and the low-frequency layer still contains pores, so smoothing does little; too large (20+) and you start softening facial contours, which reads as a plastic face. Start at 8 and move in 2-unit steps.

Third, if you're feeding it an already-upscaled 2K portrait, the same radius covers fewer feature pixels than it would at 512. Scale your radius with resolution - a rule of thumb is roughly 1 unit of radius per 100px of face size. And when the auto skin detection fails on unusual lighting or heavy makeup, feed it your own mask from a segmentation pass rather than fighting the slider.