Skip to content

Dust & Spot Removal

Automatic dust-spot cleanup, like running the Spot Healing Brush over every speck at once. It only touches small spots and leaves eyes, hair and edges alone.

Dust & Spot Removal — before / after
In ComfyUI Add Node → ComfyUI-JH-PixelPro → filters → Filters: Dust & Spot Removal
Good for Sensor dust on skies and backdrops, small specks on skin, scanned film dust, small blemishes

When to use it

Dust on the sensor shows up as small dark dots, most visible on smooth areas such as sky, studio backdrops and skin. This node looks for small spots that are darker (or brighter) than the area around them, then fills each one from the surrounding colour.

A spot only counts if it fits inside spot_size pixels. Bigger shapes such as eyes, moles you want to keep, and long thin things like hair strands are never touched.

Quick start

  1. Load Image → Filters: Dust & Spot Removal → Preview Image.
  2. Keep mode on dark for sensor dust. Use bright for white specks, both for either.
  3. Set spot_size slightly bigger than the biggest speck you want gone.
  4. Look at the spot_mask output (with Mask Preview) to see what was removed. Raise or lower sensitivity until only the specks are selected.

Settings

Setting Default What it does Tip
spot_size 12 Largest spot diameter in pixels (2 to 64). Anything bigger or longer is ignored. Keep it small on skin so freckles and pores stay.
sensitivity 0.5 How faint a spot can be and still be caught (0 to 1). Higher catches fainter spots. Raise it slowly; too high starts to pick up skin texture.
mode dark dark, bright or both. dark for sensor dust.
mask (optional) none Spots are only removed where the mask is above 0.5. Use a skin or sky mask.
protect_mask (optional) none Spots are never removed where this mask is above 0.5. Paint over freckles or beauty marks you want to keep.
fill_mode (optional) blur blur = smooth fill from the surrounding colour (the original behaviour). texture = copies the best-matching nearby patch, so skin pores, film grain or fabric continue through the spot. Use texture on skin, grain or fabric where a smooth patch would look flat. About 2 s at 4K.

fill_mode was added in v1.8.0. The default blur keeps results identical to before. Detection is the same in both modes, so spot_mask is identical.

Outputs

Output What it is
image The cleaned image.
spot_mask White (1) where pixels were replaced, black (0) elsewhere. Everything outside it is unchanged.

Tips from retouching

  • On portraits, connect a skin mask to mask (and the eyes and hair to protect_mask). The node already skips thin or elongated shapes and spots that sit in busy surroundings, so eyelashes, hair and edges are left alone, but a mask is still the safest way to keep real details such as moles or catchlights.
  • Check spot_mask first. If it shows long lines or lots of skin pores, lower sensitivity or spot_size.
  • Limit the node to the skin with mask, or protect the eyes and hair with protect_mask.
  • For heavier skin work, follow it with Skin Smoother or work on a separate layer with Frequency Separation.
  • Run it before sharpening, otherwise Smart Sharpen makes dust more visible.
  • Very big dust blobs (larger than 64 px) need manual cloning.
Under the hood

Luminance (Rec. 709 weights) is compared with a median background computed on a reduced-size copy (area average, then a 9 × 9 median in row chunks), so no big-kernel median ever runs at full resolution. The difference is thresholded (0.01 + 0.2 × (1 − sensitivity)²), then connected blobs are kept only if they are compact: bounding box at most spot_size pixels, aspect ratio at most 2, fill ratio at least 0.5 and at least about 3 px thick (a distance-transform peak of 1.3 px or more), so strands of hair and eyelashes are dropped. A blob must also sit in calm surroundings: the luminance spread in a window around it (with the blobs themselves left out) must stay below 0.25 × the spot contrast, which rejects spots on edges and in detail. On the clean sample portrait this changes about 0.006 % of the pixels. The blobs are dilated by max(1, spot_size // 6) pixels and limited by the masks. Filling uses normalised Gaussian convolution of the pixels outside the spots, computed at the reduced size and scaled back up, then blended in with a 1 px soft edge. A 2048 × 2048 image peaks at under 0.5 GB of GPU memory. If no spot is found the input is returned unchanged. With fill_mode set to texture, each spot is filled instead from the best-matching nearby patch, chosen by comparing the ring of pixels around the spot. The patch's brightness and colour are then matched to the surroundings. Source patches that contain a speck of their own are rejected, and where no clean patch fits, the spot falls back to the blur fill. Source: core/dust_spot.py.

Example workflow

Dust & Spot Removal example workflow cleans 70 synthetic sensor-dust specks from the portrait at default settings and marks what it found in red. It uses only core ComfyUI nodes and this pack. Copy workflows/sample_dust.jpg into ComfyUI/input/.

Result: Dust & Spot Removal example workflow