Image: Filters#

Blur, sharpen, edge detection, convolution, morphology.

25 nodes.

Back to the reference index

Contents#

• COP Blur (PCG) — Applies a separable Gaussian blur to the selected layers.
• COP Contour (PCG) — Draws iso-lines of a value (contour lines) at regular intervals.
• COP Convolve (PCG) — Convolves layers with a kernel written as numbers, a preset, or another image.
• COP Corner Detect (PCG) — Detects corners with the Harris detector and marks them in a mask.
• COP Defocus (PCG) — Blurs with a bokeh-shaped kernel, optionally driven by a depth layer.
• COP Denoise TV (PCG) — Removes noise with total-variation (ROF) denoising while keeping step edges.
• COP Derivative (PCG) — Computes derivatives of a layer with central differences: gradient, Laplacian, divergence or curl.
• COP Directional Blur (PCG) — Blurs along a direction by averaging samples along a line, for motion blur or light streaks.
• COP Dither (PCG) — Quantizes layers to a small number of levels using ordered, error-diffusion or random dithering.
• COP Edge Detect (PCG) — Detects edges as the gradient magnitude of a layer, optionally also outputting the x/y gradients and direction.
• COP Edge Detect Probe (PCG) — Finds edges by probing a neighbourhood for differing ids, depth jumps or normal creases, in the manner of Houdini's Edge Detect by Contour, Depth and Normal.
• COP Erode / Dilate (PCG) — Applies morphological operations (erode, dilate, open, close, gradient) with a square or round window.
• COP Extrapolate Boundaries (PCG) — Extends the valid interior of an image outward by copying nearest interior values, mainly to bleed texture edges.
• COP Feather (PCG) — Softens the edge of a mask with a smoothstep ramp based on the exact distance to the boundary.
• COP Feather Slope (PCG) — Limits the maximum rate of change between neighbouring pixels by lifting dark pixels toward bright ones (or the reverse).
• COP FFT (PCG) — Performs a forward or inverse Fourier transform, shows the magnitude spectrum, or filters frequencies.
• COP Glow (PCG) — Adds a glow or star-shaped light streaks around the bright parts of an image.
• COP Kuwahara (PCG) — Smooths an image while keeping edges, giving a painterly look.
• COP Median (PCG) — Applies a median filter that removes speckle while preserving edges.
• COP Pixelate (PCG) — Pixelates an image into square blocks.
• COP SDF from Mask (PCG) — Converts a mask into a signed distance field, negative inside and positive outside.
• COP SDF Ops (PCG) — Combines, offsets or converts signed distance fields.
• COP Sharpen (PCG) — Sharpens with an unsharp mask.
• COP Smooth Fill (PCG) — Fills unknown pixels with smooth values interpolated from the known ones.
• COP Surface Dither (PCG) — Draws halftone dots attached to a UV parametrisation whose spacing stays near a fixed pixel size under scaling and perspective.

Nodes#

COP Blur (PCG)#

PCG_COPBlur · category PCG/cop · Houdini: Blur (COP)

Applies a separable Gaussian blur to the selected layers.

The blur is separable: a horizontal pass followed by a vertical pass, so the cost grows with the kernel width rather than its square. The kernel radius is ceil(3*sigma), which keeps about 99.7% of the Gaussian energy; the truncated kernel is renormalised. Cutting at 2 sigma would lose about 4.6% and leave a visible hard rim on large soft glows.

A blur should preserve the average brightness, but only with the wrap border. clamp copies the edge pixels and black pads with zero, so both invent data outside the image and the mean shifts; this is the definition of those borders, not an error.

With Layers empty every layer is blurred. List layer names to blur colour but leave, for example, an SDF layer alone.

Inputs

Outputs

Parameters

Tips

• Use wrap when blurring a tileable texture so the seams stay seamless.

COP Contour (PCG)#

PCG_COPContour · category PCG/cop

Draws iso-lines of a value (contour lines) at regular intervals.

A line is drawn wherever the value of Layer crosses a multiple of Interval (shifted by Offset). The distance to the nearest line is measured in value units and divided by the local gradient magnitude, which converts it to a distance in pixels. As a result the line width is the same in pixels on steep and on flat areas; without this step lines on gentle slopes would smear into broad bands. The output is a single-channel layer between 0 and 1.

Inputs

Outputs

Parameters

Compared with Houdini

Corresponds to the contour mode of Houdini's Edge Detect.

COP Convolve (PCG)#

PCG_COPConvolve · category PCG/cop · Houdini: Convolve 3x3 (COP), Convolve Filter (COP)

Convolves layers with a kernel written as numbers, a preset, or another image.

The kernel is n x n numbers (n odd), written one row per line. The first line written is the top row of the image. Presets: box, gaussian, sharpen, edge, emboss and Sobel X / Y. Kernels larger than 7 x 7 are computed with an FFT.

When the kernel_cop input is connected, a layer of that image (Kernel Layer) is used as the kernel instead. This is a true convolution: convolving a single-pixel impulse gives the kernel image itself. Preset and Kernel are then ignored. With Centered off, the origin is the bottom-left pixel of the kernel image.

The image is always padded by the kernel radius, so no wrap-around bleeding occurs with clamp, black or mirror.

Inputs

Outputs

Parameters

Compared with Houdini

Covers Convolve 3x3 and Convolve Filter. Houdini's Padding option is not exposed: the image is always padded by the kernel radius (equivalent to Houdini's Padding on and large enough), so there is no wrap-around bleeding that Houdini shows with Padding off. Border has an extra mirror option.

COP Corner Detect (PCG)#

PCG_COPCornerDetect · category PCG/cop · Houdini: Corner Detect (COP)

Detects corners with the Harris detector and marks them in a mask.

Computes the structure tensor of the first channel of Layer, smoothed with a Gaussian of Sigma, and the Harris response R = det - k * trace^2. A pixel is marked 1 when its response is greater than Threshold times the maximum response and it is the local maximum within Radius pixels. The second output is the number of corners found.

Inputs

Outputs

Parameters

COP Defocus (PCG)#

PCG_COPDefocus · category PCG/cop · Houdini: Bokeh (COP), Defocus (COP)

Blurs with a bokeh-shaped kernel, optionally driven by a depth layer.

The image is convolved with an aperture-shaped kernel (disc, polygon or square) of Radius pixels. Without a Depth Layer the whole image uses the same radius.

With a Depth Layer the radius varies per pixel: pixels where |depth - Focus| <= Focus Range stay sharp, and beyond that the circle of confusion grows linearly over the Falloff distance up to Radius. The radius range is quantized into Levels steps; each step is convolved once and neighbouring steps are interpolated per pixel. The depth layer itself is not blurred.

Highlight Gain makes bright areas stand out: values above Highlight Threshold are boosted before blurring and mapped back with the exact inverse afterwards, so bright spots become distinct aperture-shaped discs instead of being averaged into grey.

Inputs

Outputs

Parameters

Depth

Highlights

Compared with Houdini

Covers Defocus and Bokeh.

COP Denoise TV (PCG)#

PCG_COPDenoiseTV · category PCG/cop · Houdini: Denoise TVD (COP)

Removes noise with total-variation (ROF) denoising while keeping step edges.

Solves the Rudin-Osher-Fatemi total-variation model with Chambolle's dual projection. Noise is flattened while sharp edges are preserved. Each channel is processed independently. Weight sets the strength: higher is flatter. The mean of the image is preserved exactly and a constant image is returned unchanged.

Inputs

Outputs

Parameters

Compared with Houdini

Corresponds to Denoise TVD.

COP Derivative (PCG)#

PCG_COPDerivative · category PCG/cop · Houdini: Curl (COP), Derivative (COP), Laplacian (COP)

Computes derivatives of a layer with central differences: gradient, Laplacian, divergence or curl.

Derivatives are central differences in units of per pixel, multiplied by Scale.

• dx, dy, gradient (two channels) and magnitude use the first channel.
• laplacian uses a five-point stencil per channel.
• divergence turns a two-channel vector field into a scalar.
• curl on a one-channel layer gives (df/dy, -df/dx), a divergence-free velocity field, which is how curl noise is made from a noise stream function. On a two-channel layer it gives the scalar curl dv/dx - du/dy.

If Out Layer is empty, the result is named layer_mode.

Inputs

Outputs

Parameters

Compared with Houdini

Covers Derivative, Laplacian and Curl.

COP Directional Blur (PCG)#

PCG_COPDirectionalBlur · category PCG/cop · Houdini: Streak Blur (COP)

Blurs along a direction by averaging samples along a line, for motion blur or light streaks.

Each pixel averages bilinear samples along a line of Length pixels in the direction given by Angle (0 degrees points right, 90 degrees up). In centered mode the samples cover half the length on either side (motion blur). In streak mode the samples run from the pixel back along the line, so content smears one way, in the direction of Angle (light streaks, speed lines). A negative length reverses the direction. With Samples at 0 the node takes about one sample per pixel of length.

Inputs

Outputs

Parameters

Compared with Houdini

Corresponds to Streak Blur and can also be used as motion blur.

COP Dither (PCG)#

PCG_COPDither · category PCG/cop

Quantizes layers to a small number of levels using ordered, error-diffusion or random dithering.

Equivalent of Surface Dither. Values are first clamped to 0-1, then reduced to levels evenly spaced steps with a dithering pattern that preserves the average brightness.

Bayer methods are ordered dithering: the pattern tiles and does not flicker in animation. Floyd-Steinberg is error diffusion, processed serially (scanline order, top to bottom), and is the closest to the original average brightness but slower on large images. White noise uses a random threshold per pixel.

Inputs

Outputs

Parameters

COP Edge Detect (PCG)#

PCG_COPEdge · category PCG/cop · Houdini: Edge Detect (COP)

Detects edges as the gradient magnitude of a layer, optionally also outputting the x/y gradients and direction.

The main output is the gradient magnitude. With components on, three more layers <out>_gx, <out>_gy and <out>_dir (direction angle) are written as well.

Kernels are normalized to true derivatives. Raw integer Sobel kernels return 8 times the real gradient (Prewitt 6, Scharr 32). Left alone, switching kernels would change edge strength several-fold and every threshold would need retuning. Here every kernel is divided out, so on a linear ramp f = a*x + b*y the gx and gy outputs are exactly a and b. The consequence is that values look smaller than in tools using raw integer kernels. The Laplacian is not normalized; it is the plain second derivative sum, so f = a*x^2 gives 2a.

Multi-channel layers are processed per channel, and the magnitude is the root of the summed squares, not a grayscale conversion first. A red-green boundary with equal luminance would vanish after conversion. If you want luminance-based edges, put a COP Luminance in front.

Inputs

Outputs

Parameters

Tips

• Edge strength is a true per-pixel gradient, so a unit-range image rarely exceeds 1. Raise scale before thresholding for display.

COP Edge Detect Probe (PCG)#

PCG_COPEdgeDetectProbe · category PCG/cop · Houdini: Edge Detect by Contour (COP), Edge Detect by Depth (COP), Edge Detect by Normal (COP)

Finds edges by probing a neighbourhood for differing ids, depth jumps or normal creases, in the manner of Houdini's Edge Detect by Contour, Depth and Normal.

For every pixel, integer-offset neighbours inside a probe radius are examined. The radius is thickness_scale (multiplied by the optional thickness layer), not smaller than min_probe_radius, and capped at 64 pixels. A neighbour is on the line if its value differs from the pixel by more than contour_tolerance (Contour, Euclidean distance over channels, normally an id layer), if its depth differs by more than depth_tolerance in either direction (Depth), or if the dot product of the normalized normals is below normal_tolerance (Normal). The pixel's result is the strongest weight among such neighbours: 1 for immediate neighbours, falling off as a Gaussian of the distance with sigma = edge_spread times the radius, cut off at the radius.

This is separate from COP Contour (iso-lines) and from the lineart ControlNet packaging in Control Image. Houdini's documentation does not specify the algorithm in detail; this is the plugin's own definition. There is no camera, so tolerances are absolute values, and pixels outside the image repeat the edge pixel.

Inputs

Outputs

Parameters

Compared with Houdini

Combines Edge Detect by Contour, by Depth and by Normal into one node selected by Kind. Without a camera or ray differentials, tolerances are absolute and not scaled by distance or slope.

COP Erode / Dilate (PCG)#

PCG_COPMorph · category PCG/cop · Houdini: Dilate Erode (COP)

Applies morphological operations (erode, dilate, open, close, gradient) with a square or round window.

Erode takes the local minimum and dilate the local maximum over the window, in pixels. On masks this shrinks or grows shapes; on height fields it flattens peaks or fills valleys. Open (erode then dilate) removes small specks, Close (dilate then erode) fills small holes, and Gradient (dilate - erode) gives an outline about twice the radius thick on binary masks.

The square window is separable (two 1D passes, cost proportional to the radius), the round one is not and costs one full-image minimum per offset in the window (197 at radius 8, 1257 at radius 20). For large radii use an SDF instead: COP SDF from Mask followed by subtracting the radius gives a dilation of any size at a cost that does not depend on the radius, and is identical to a disc dilation of a binary mask. Use this node for small fixes (radius up to about 8).

Border matters: Clamp copies the edge, so the image border does not take part in erosion; Constant (Black) treats the outside as 0, so erosion eats inward from all four edges.

Inputs

Outputs

Parameters

Compared with Houdini

Corresponds to Houdini's Dilate Erode, extended with Open, Close and Gradient.

COP Extrapolate Boundaries (PCG)#

PCG_COPExtrapolateBoundaries · category PCG/cop · Houdini: Extrapolate Boundaries (COP)

Extends the valid interior of an image outward by copying nearest interior values, mainly to bleed texture edges.

The fill-area layer is split into inside and outside by threshold and inner_side. Outside pixels within Edge Padding (in image units) copy the nearest interior value; this can be pulled inward by Edge Offset and can restrict sampling to interior pixels only. Pixels beyond the padding are handled by exterior: left unchanged, set to black, or filled by diffusion.

This is distinct from COP Smooth Fill, which fills holes in a mask with three different algorithms.

Interior pixels are never changed. Errors if the fill-area layer has no interior pixels. A value exactly equal to the threshold counts as exterior.

Inputs

Outputs

Parameters

COP Feather (PCG)#

PCG_COPFeather · category PCG/cop

Softens the edge of a mask with a smoothstep ramp based on the exact distance to the boundary.

The mask is thresholded at 0.5 and the ramp of width width pixels follows the exact distance to the boundary. Unlike a Gaussian blur, thin strips do not fade overall and sharp corners are not rounded off. In both mode the boundary itself is always exactly 0.5.

In inward mode strips narrower than the width do get lighter overall; that is the literal meaning of softening inward only.

Inputs

Outputs

Parameters

COP Feather Slope (PCG)#

PCG_COPFeatherSlope · category PCG/cop · Houdini: Feather (COP)

Limits the maximum rate of change between neighbouring pixels by lifting dark pixels toward bright ones (or the reverse).

Equivalent of Houdini's Feather. With High to Low on, bright pixels stay fixed and the darker surroundings are raised until the slope limit is satisfied; with it off, dark pixels stay and bright ones are pushed down. The slope is given by the Feather Method (distance per unit change, or change per distance, in image units). The shape of the resulting islands comes from the sweep directions selected by direction (Diamond, Square, Octagon, Custom, Circle).

This differs from the older COP Feather, which applies a smoothstep by distance to the mask boundary.

Inputs

Outputs

Parameters

COP FFT (PCG)#

PCG_COPFFT · category PCG/cop · Houdini: Fast Fourier Transform (COP), Filter Frequencies (COP)

Performs a forward or inverse Fourier transform, shows the magnitude spectrum, or filters frequencies.

Covers Houdini's Fast Fourier Transform and Filter Frequencies. forward turns channel 0 into a 2-channel layer (real, imaginary) without normalization (the DC component is the sum of the pixels); inverse takes (real, imaginary) back to one real channel; magnitude outputs log(1 + |F|) per channel for viewing the spectrum; filter transforms, multiplies by a radial window and transforms back, per channel, with a cosine soft edge (a hard cutoff would ring).

Cutoffs are in cycles per image (image width = 1 cycle); the radial distance is elliptical, each axis scaled by its own size.

Inputs

Outputs

Parameters

Filter

COP Glow (PCG)#

PCG_COPGlow · category PCG/cop · Houdini: Glow (COP), Star Glow (COP)

Adds a glow or star-shaped light streaks around the bright parts of an image.

Covers Glow and Star Glow. Pixels brighter than threshold (with a soft knee, using the maximum of the first three channels) are extracted. In glow mode the extracted light is blurred at three scales (radius, half and quarter radius) and averaged; in star mode it is smeared along rays evenly spaced directions. The result times intensity is added back to the original. Alpha is left unchanged.

Inputs

Outputs

Parameters

Star

COP Kuwahara (PCG)#

PCG_COPKuwahara · category PCG/cop · Houdini: Kuwahara Filter (COP)

Smooths an image while keeping edges, giving a painterly look.

The classic Kuwahara filter: around each pixel the neighbourhood is split into four quadrants, and the pixel takes the mean of the quadrant with the lowest variance. Flat areas are smoothed, edges stay sharp, and larger radii give a stronger oil-paint look.

Inputs

Outputs

Parameters

COP Median (PCG)#

PCG_COPMedian · category PCG/cop · Houdini: Median (COP)

Applies a median filter that removes speckle while preserving edges.

Each pixel takes the median of its neighbourhood. This removes salt-and-pepper noise and small blotches on AI-generated masks without blurring edges. A constant image is unchanged, and every output value is one of the values present in the window.

The filter needs sorting, so it runs on the CPU (a GPU image makes a round trip) and costs roughly O(N*k^2*log k): radius 2 is already visibly slower than a blur and radius 4 much slower. Do not use it as a blur; the radius is limited to 8.

Inputs

Outputs

Parameters

COP Pixelate (PCG)#

PCG_COPPixelate · category PCG/cop · Houdini: Pixelate (COP)

Pixelates an image into square blocks.

Divides the image into Size x Size pixel blocks and fills each block with its average color or with the value at the block centre.

Inputs

Outputs

Parameters

COP SDF from Mask (PCG)#

PCG_COPSDFFromMask · category PCG/cop · Houdini: Mono to SDF (COP)

Converts a mask into a signed distance field, negative inside and positive outside.

Distances are in pixels or in UV units (divided by the short edge, the same convention as COP Shape, so a baked Shape circle gives comparable values). The sign convention matches COP Shape. An SDF enables operations a black and white mask cannot: outlines (|d| < w) with sub-pixel exact, resolution-independent thickness; grow/shrink by d - r with r in pixels; feathering by smoothstep(-w, w, d); union/intersect/subtract by min/max, which keep a valid distance field; and equidistant bands via Quantize.

Exact uses the Felzenszwalb lower-envelope algorithm (two 1D passes), which gives the true Euclidean distance; it is NumPy only. Jump Flooding is gather only and can run on GPU, but is approximate (a solid shape retains a ~0.19 px error that extra passes do not remove). On the CPU Exact is both faster and more accurate, so there is no reason to use JFA there.

Half-Pixel Correction: distances are measured between pixel centres, so a pixel touching the boundary reads 1.0 while the true boundary is at 0.5. Subtracting 0.5 fixes this near the boundary. It does not move the zero level set, so thresholding gives identical results either way; it only changes the numeric values.

Inputs

Outputs

Parameters

COP SDF Ops (PCG)#

PCG_COPSDFOps · category PCG/cop · Houdini: SDF Blend (COP), SDF to Mono (COP)

Combines, offsets or converts signed distance fields.

Union, Intersect and Subtract combine SDF A with SDF B; with a Smooth Radius above 0 a polynomial smooth version gives rounded transitions. Blend morphs between the two shapes. Offset grows (positive) or shrinks the shape and Onion turns it into a shell. The SDF to Mono operations turn the field into a mask: filled shape, outline, or falloff, with Width in the chosen Units.

Inputs

Outputs

Parameters

COP Sharpen (PCG)#

PCG_COPSharpen · category PCG/cop · Houdini: Sharpen (COP)

Sharpens with an unsharp mask.

Computes out = a + Amount * (a - blur(a, Sigma)). Amount 0 is a bit-exact identity (the node is skipped), and a constant image is unchanged for any setting.

Sigma sets the scale of the detail being enhanced: 0.5 to 1 brings out texture, 3 to 5 brings out contours and volume. Amounts above 2 produce halos (overshoot) on strong edges; this is inherent in the method.

Inputs

Outputs

Parameters

COP Smooth Fill (PCG)#

PCG_COPSmoothFill · category PCG/cop · Houdini: Smooth Fill (COP)

Fills unknown pixels with smooth values interpolated from the known ones.

Pixels where the Known Mask Layer is above 0.5 are known; all other pixels are filled. Push-Pull uses an image pyramid and is fast. Laplace solves a harmonic membrane and is the smoothest. Nearest extrapolates ring by ring. Known pixels are left bit-exact.

Inputs

Outputs

Parameters

COP Surface Dither (PCG)#

PCG_COPSurfaceDither · category PCG/cop · Houdini: Surface Dither (COP)

Draws halftone dots attached to a UV parametrisation whose spacing stays near a fixed pixel size under scaling and perspective.

Surface Dither lays a grid of dots in pattern space (the UVs, scaled, rotated and offset) and picks a grid level per pixel from the screen-space derivative of that pattern space, so the on-screen dot spacing stays between Density Scale / 2 and Density Scale pixels. When the level changes, the dots that must disappear (three out of four) shrink away during the last Level Transition fraction of the level, so zooming or foreshortening neither blurs the dots into grey nor makes them pop.

The result is either a coverage mask or a signed distance field (in pixels), written to Output Layer. Optionally it also writes the local cell UV, the level index and the relative dot scale as extra layers. If no UV layer is given, the texture UV of the pixel centre is used. Where the UVs have a seam (a jump), the derivative is huge and a very coarse level is chosen there.

This is a UV-attached halftone, not the screen-space quantisation dither (Bayer / error diffusion) of the plain Dither node.

Inputs

Outputs

Parameters

Compared with Houdini

Custom dot shapes (the shape input and Shape Resolution) are not implemented; dots are always circles. Houdini does not document how seams in the UVs are handled, so the behaviour there is simply the derivative-based level choice described above.

Tips

• Feed a perspective-projected UV layer (for example from a rasterised mesh) to see the level selection at work.
• Output a scale_layer and use it to tint or debug where the level changes.