Image: Channels & Layers#

Channels, layers, conversions between data types.

13 nodes.

Back to the reference index

Contents#

• Control Image (PCG) — Packs a COP layer into a ComfyUI IMAGE and MASK following the conventions of ControlNet inputs.
• COP Channel Convert (PCG) — Converts a layer between channel layouts such as mono, RGB, RGBA and UV.
• COP Channel Join (PCG) — Joins all channels of several layers into one layer of at most four channels.
• COP Channel Split (PCG) — Splits a layer into one single-channel layer per channel.
• COP Convert Depth (PCG) — Converts between depth representations: camera z, distance to camera, disparity, normalized depth and camera-space position.
• COP Convert Normal (PCG) — Converts normal maps between packed/signed encoding and OpenGL/DirectX conventions.
• COP ID Convert (PCG) — Converts an integer ID layer into a mask, a normalized mono value or a signed distance field.
• COP Layer Edit (PCG) — Renames, copies, deletes or keeps layers of an image.
• COP Layer Merge (PCG) — Copies layers from image B into image A.
• COP Luminance (PCG) — Converts a layer to a single channel (luminance, average, HSV component, one channel, or custom weights).
• COP Swizzle (PCG) — Rebuilds a layer from a pattern of channel letters and constants.
• ID Masks (PCG) — Splits an ID layer into a batch of masks, one per region.
• Regional Prompt from COP (PCG) — Builds region-based conditioning from COP mask layers and one prompt per region.

Nodes#

Control Image (PCG)#

PCG_ControlImage · category PCG/cop · Houdini: ID to RGB (COP)

Packs a COP layer into a ComfyUI IMAGE and MASK following the conventions of ControlNet inputs.

Converts a layer to the exact value convention a ControlNet expects. The conversion itself is mechanical; the conventions are what make it useful, because a wrong depth direction or an extra transfer curve gives images that ignore the control without any error. Per kind:

• Depth: inverse depth, near = white, as produced by MiDaS-style estimators (renderers give camera-space z where far is large, so it has to be reversed). By default the range is the min/max of the pixels inside the mask (relative depth, fits the object); fill Near / Far to keep it stable across an animation. Background is 0 (black = farthest); a value of 1 would make the model think something touches the lens.
• Height: top-down depth of a height field, high = white. Do not feed height to Depth, which reverses it (mountains become pits). Near / Far are the height range here; 0 = automatic. For a top-down view use the Normal kind with the tangent-space N from Height to Normal.
• Normal: packs n x 0.5 + 0.5, background (0.5, 0.5, 1). The normals must be camera-space (Render Pass default). The green channel convention differs between OpenGL (+Y up) and DirectX (+Y down) and between ControlNets; use the flip switches. A wrong choice shows as inverted bumps rather than a failure.
• Line Art: extracted from discontinuities of depth and normals instead of image edge detection: the mask gradient gives the outer silhouette, the depth gradient inner occlusion edges, the normal gradient creases. Cleaner than rendering and running Canny, which also picks up materials, noise and shading gradients, and silhouette and crease can be weighted separately.
• Segmentation (ID): each ID gets a distinct color via a golden-angle hue sequence (random hues collide, equal steps need the count in advance).
• Mask, Color, Raw: the layer as grey, color, or unchanged.

COP values are linear while a VAE sees sRGB-like 0-1 values. Depth, normal, line art and ID are data and ControlNets are trained in the packed space, so no curve is applied; Color for img2img needs the sRGB curve. Image and mask are resized to the nearest multiple of Align (the latent is 1/8 resolution, so other sizes are silently rescaled somewhere), moving at most half the multiple. The output is flipped vertically, because COP row 0 is the bottom and IMAGE row 0 is the top.

Inputs

Outputs

Parameters

Depth

Normal

Line Art

ID

Compared with Houdini

Its ID kind corresponds to Houdini's ID to RGB.

Tips

• Fix Near and Far for animations, otherwise the depth range changes from frame to frame.

COP Channel Convert (PCG)#

PCG_COPChannelConvert · category PCG/cop · Houdini: Mono to RGBA (COP), RGB to RGBA (COP), RGB to UV (COP), RGBA to RGB (COP), RGBA to UV (COP), UV to RGB (COP), UV to RGBA (COP)

Converts a layer between channel layouts such as mono, RGB, RGBA and UV.

One menu replaces Houdini's family of conversion nodes: mono to RGB / RGBA, RGB to RGBA, RGBA to RGB, RGB / RGBA to UV, UV to RGB / RGBA, and RGB to mono by luma, average or maximum.

If the source layer has fewer channels than the conversion needs, the node reports an error instead of silently padding with zeros, so a wrong connection (for example RGB treated as UV) does not travel downstream unnoticed. Extra channels are truncated. Channels that have to be created get Alpha when it is the alpha channel and Fill otherwise.

Inputs

Outputs

Parameters

Compared with Houdini

Covers Mono to RGBA, RGB to RGBA, RGBA to RGB, RGB to UV, RGBA to UV, UV to RGB and UV to RGBA. For luminance from RGB there is also COP Luminance.

COP Channel Join (PCG)#

PCG_COPChannelJoin · category PCG/cop · Houdini: Channel Join (COP)

Joins all channels of several layers into one layer of at most four channels.

Takes the layers listed in Layers in order and concatenates all their channels into a single layer. The usual use is joining three single-channel layers into an RGB layer. The total number of channels must not exceed four.

Inputs

Outputs

Parameters

COP Channel Split (PCG)#

PCG_COPChannelSplit · category PCG/cop · Houdini: Channel Split (COP)

Splits a layer into one single-channel layer per channel.

Each channel of Layer becomes its own layer named layer.suffix, for example C.r, C.g, C.b, C.a. Suffixes are taken in order from Suffixes; if there are more channels than suffixes the channel index is used.

Inputs

Outputs

Parameters

COP Convert Depth (PCG)#

PCG_COPConvertDepth · category PCG/cop · Houdini: Convert Depth (COP)

Converts between depth representations: camera z, distance to camera, disparity, normalized depth and camera-space position.

Renderers usually give z depth along the view axis, ray tracers give the Euclidean distance to the camera, and depth ControlNets or monocular depth estimators use inverse depth (disparity). These differ by a per-pixel factor of sqrt(1 + x^2 + y^2), which is more than 20% at the image edge for a 60 degree field of view. Connecting the wrong kind produces no error: objects near the edges are just pushed too far away.

z_depth is the camera-space z, distance the distance to the camera, disparity is 1/z, normalized is (z - near) / (far - near) with 0 at the near plane, and position (output only) is the camera-space point (x, y, z). The field of view is the vertical one, as in the Render Pass convention. Pixels outside the mask, or with a value of 0 or less when no mask layer is given, are written as 0.

Inputs

Outputs

Parameters

Tips

• Check which kind of depth your source delivers before using it with a depth ControlNet; z depth and distance differ visibly near the image edge.

COP Convert Normal (PCG)#

PCG_COPConvertNormal · category PCG/cop · Houdini: Convert Normal (COP)

Converts normal maps between packed/signed encoding and OpenGL/DirectX conventions.

Steps always run in this fixed order: decode, reconstruct z, flip, normalize, encode.

• packed vs signed: image files store normals as 0..1 colours, while inside this plugin normals are vectors in -1..1. A normal map read in from ComfyUI must go through packed to signed first.
• Flip Green switches between OpenGL (+Y up, used by this plugin) and DirectX (+Y down). Getting this wrong inverts bumps along one direction and raises no error.
• Reconstruct Z rebuilds z as sqrt(1 - x^2 - y^2) from the x and y channels only (two-channel BC5-style normal maps).

packed to signed to packed is an identity, and flipping twice restores the input exactly. Conversion between tangent space and world space is not supported because it needs a tangent frame per pixel, which the height-map chain does not have.

Inputs

Outputs

Parameters

COP ID Convert (PCG)#

PCG_COPIDConvert · category PCG/cop · Houdini: ID to Mask (COP), ID to Mono (COP), ID to SDF (COP)

Converts an integer ID layer into a mask, a normalized mono value or a signed distance field.

Reads the first channel of the ID layer, rounded to integers.

• Mask: IDs listed in IDs become 1, everything else 0.
• Mono: the ID value itself. With Normalize it is mapped to 0-1 as (id - min) / (max - min).
• SDF: with IDs given, the signed distance (negative inside, positive outside, boundary at plus or minus half a pixel) of the union of the selected IDs; Invert uses the complement. With IDs empty (or *), each pixel gets its distance to the nearest pixel with a different ID, so the per-region interior distances are combined in one image. The SDF mode fails if the selection is empty or covers the whole image, because there is then no boundary.

Inputs

Outputs

Parameters

Compared with Houdini

One node covers Houdini's ID to Mask, ID to Mono and ID to SDF, selected with Convert To.

COP Layer Edit (PCG)#

PCG_COPLayerEdit · category PCG/cop · Houdini: Rename (COP)

Renames, copies, deletes or keeps layers of an image.

Edits the layer set of an image without touching pixel data. Renaming onto an existing name overwrites that layer. Keep Only fails if no layer would remain. Rename and Copy take space-separated old=new pairs; Delete and Keep Only take a list of layer names.

Inputs

Outputs

Parameters

Compared with Houdini

Covers Houdini's Rename COP and adds copy, delete and keep operations in the same node.

COP Layer Merge (PCG)#

PCG_COPLayerMerge · category PCG/cop

Copies layers from image B into image A.

Adds the layers of B to A without doing any pixel math. If the resolutions differ, B is resampled to A's resolution; if the backends differ, B is moved to A's backend. Prefix is added to the incoming layer names, and with Overwrite Existing turned off, layers of A that have the same name are kept.

To combine the pixels of the same layer from several images, use COP Layer Combine instead.

Inputs

Outputs

Parameters

COP Luminance (PCG)#

PCG_COPLuminance · category PCG/cop · Houdini: HeightField to Mono (COP), Mono (COP)

Converts a layer to a single channel (luminance, average, HSV component, one channel, or custom weights).

Reduces a layer to one channel stored as a new layer. The default HDTV operation is Rec.709 luminance (0.2126, 0.7152, 0.0722). COP data is linear, so weighting directly is correct; applying the same weights to sRGB-encoded values is a common mistake (decode first).

Other operations: SGI and NTSC use their own weights; Average, Maximum Channel and Minimum Channel use all existing channels (alpha included); Magnitude is the vector length (absolute value for one channel); Hue, Saturation and Value convert to HSV (hue in 0-1, not degrees); Red/Green/Blue/Alpha pick one channel, with missing channels filled by the usual expansion rules; Custom computes the weighted sum of the channels, optionally normalizing the weights in use to sum to 1.

Layers with fewer than three channels use their first channel instead of raising an error, so masks and SDFs can be passed through.

Inputs

Outputs

Parameters

Compared with Houdini

Corresponds to Houdini's Mono COP and its Operation menu; also covers HeightField to Mono.

COP Swizzle (PCG)#

PCG_COPSwizzle · category PCG/cop · Houdini: Channel Extract (COP), Channel Swap (COP)

Rebuilds a layer from a pattern of channel letters and constants.

Each character of Pattern produces one output channel, so the output has as many channels as the pattern has characters (1 to 4). Letters pick a channel of the source layer: r g b a, x y z w or u v. The characters 0 and 1 produce constant channels.

Examples: bgr swaps red and blue, a extracts alpha as a single-channel layer, rrr1 turns a grey image into RGBA. A request for a channel the source layer does not have (for example a on a 3-channel layer) is an error.

Inputs

Outputs

Parameters

Compared with Houdini

Combines Houdini's Channel Swap and Channel Extract in one pattern string.

Tips

• Use a with a different output layer to pull alpha out as a mask.

ID Masks (PCG)#

PCG_IDMasks · category PCG/cop

Splits an ID layer into a batch of masks, one per region.

ComfyUI MASKs are (B, H, W), so one mask per region is naturally a batch. Connecting it to ConditioningSetMask gives per-region prompts with geometrically exact borders instead of painted ones.

The IDs come from the id pass of Render Pass, which reads a primitive attribute such as name, piece or cluster (written by Voronoi Fracture, Voronoi Cluster, Convex Decompose, Connectivity). IDs are rounded to integers and sorted ascending. If there are more regions than Max Regions, the ones with the largest area are kept (not the highest IDs, which would cut away the main subject). Only pixels inside the mask layer are considered.

Inputs

Outputs

Parameters

Regional Prompt from COP (PCG)#

PCG_RegionalPrompt · category PCG/cop

Builds region-based conditioning from COP mask layers and one prompt per region.

Replaces N CLIP Text Encode nodes, N Conditioning Set Mask nodes and N-1 Conditioning Combine nodes with one node. Regions has one line per region in the form layer name: prompt.

In Normalized mode the layers are already a partition (for example the output of Splat Map); they are used as they are and normalised only where they add up to more than 1. In Layered mode they are independent masks (several Mask by Feature, say), composited in written order so later lines overwrite earlier ones. The base prompt by default covers only what remains (Remainder), so all masks sum to exactly 1 everywhere.

This matters because ComfyUI averages masked conditioning weighted by the mask values. A pixel not covered by any condition has a prediction of 0, which shows as a blurry hole without any error. A base prompt without mask (Global) takes part everywhere and halves the strength of the regional prompts.

Masks are composited in COP orientation and flipped vertically at the output (the same boundary as Control Image), so they align pixel for pixel with depth or normal control images from the same COP. Regions that are empty are skipped. The report output lists the coverage per region.

Inputs

Outputs

Parameters