Skip to content

Detail Mask

Finds where your photo has fine detail (eyelashes, hair, fabric weave, edges) and returns it as a mask. Use it to sharpen only where there is texture, or to keep smoothing away from detail.

Detail Mask — before / after
In ComfyUI Add Node → ComfyUI-JH-PixelPro → mask → High-Frequency Detail Masker
Good for Sharpening only textured areas, protecting detail during skin smoothing, building a mask for noise or grain work

When to use it

In frequency terms, high-frequency content is the fine, fast-changing part of a photo: pores, hair strands, edges. Low-frequency content is the smooth colour and tone. This node measures how much fine detail each pixel has, then keeps the top share of the strongest pixels.

Typical uses:

  • Gate Smart Sharpen so flat skin and sky do not get sharpened.
  • Protect eyes, lashes and hair when you run Skin Smoother.
  • Invert it to get the smooth areas only.

Quick start

  1. Connect your photo to image.
  2. Start with kernel_type laplacian and sensitivity 0.4.
  3. Preview the mask with Mask Preview. Raise sensitivity to include more.
  4. Use the mask as the gate in your sharpen or smooth step.

Settings

Setting Default What it does Tip
kernel_type laplacian How detail is measured. laplacian finds fine detail in all directions; sobel favours edges; fs_gaussian uses the difference between the photo and a light blur (the high-pass of frequency separation). Start with laplacian.
sensitivity 0.5 Share of pixels kept as detail. 0 gives an empty mask, 1 gives a full mask. 0 to 1. For portraits, 0.3 to 0.6 is typical.
threshold_mode relative_percentile relative_percentile adapts to each image, so the same sensitivity selects a similar share of pixels. absolute is scaled by the strongest detail in the image. Use the default. absolute can be thrown off by one bright speck.
mask (optional) none Limits the result to a region. The output is multiplied by it. Use a face or subject mask.

Output: mask_detail (MASK), values 0 or 1.

Tips from retouching

  • The output is a hard on/off mask. Soften it with Mask Edge Smoother or a blur before using it as a blend mask.
  • Detail is measured on the average of the colour channels (grey), so colour-only edges are weak.
  • Noisy photos look "full of detail". Denoise first if the mask lights up on grain.
  • sensitivity is a share of the image, not a strength. A flat photo still returns the requested share, just on its relatively sharpest parts.
Under the hood

Luma is the mean of the RGB channels. High-pass options: laplacian (Kornia, 3x3, replicate border), sobel (Kornia, normalised gradient magnitude) or fs_gaussian (luma minus a 5x5, sigma 1 Gaussian blur). In relative_percentile mode the threshold is the per-image k-th value of the absolute high-pass so that about sensitivity of the pixels pass. In absolute mode the high-pass is divided by its per-image maximum and compared with 1 - sensitivity. Sensitivity 0 and 1 are shortcuts for an empty or full mask. Source: core/detail_masker.py.

Example workflow

Detail Mask example workflow. Copy workflows/sample_portrait.jpg into ComfyUI/input/ first.