OpenCV Denoise (Luma/Chroma):
OpenCVDenoiseColored reduces visible grain and colored speckles in an image using OpenCV’s fast non-local means denoising method. The method compares small areas of an image with other areas that look similar, helping smooth noise while retaining more recognizable edges and texture than simple blurring. It treats brightness noise and color noise separately, so you can address one without applying the same strength to the other. Stronger denoising can also soften fine details, making a balanced setting important for artwork with delicate textures. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai))
OpenCV Denoise (Luma/Chroma) Input Parameters:
image
The IMAGE to denoise. The node processes each image in the incoming batch separately and returns them together in the same order. It expects ordinary three-channel RGB image data in the format used by the image workflow; the source converts each image to an 8-bit OpenCV image before processing. There is no minimum, maximum, or default value for this required input. An empty batch or an image without three color channels cannot be processed by this implementation.
h_luminance
A FLOAT controlling how strongly the method reduces brightness-related noise, such as light and dark grain. Its default is 3.0, its minimum is 0.0, its maximum is 100.0, and its interface step is 0.5. Raising it can clean up more grain but may also remove subtle shading and fine detail; lowering it preserves more detail but may leave noise visible. Start near the default and increase it gradually while checking textured areas and edges. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai))
h_color
A FLOAT controlling the strength of color-noise reduction, particularly small, unwanted colored speckles. Its default is 3.0, its minimum is 0.0, its maximum is 100.0, and its interface step is 0.5. Increase it when color speckling remains after brightness grain is acceptable; excessive strength can reduce subtle color variation. This setting acts on color information separately from h_luminance, so the two strengths need not match. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai))
search_window
An INT specifying the width, in pixels, of the area searched for similar image patches. Its default is 5, its minimum is 5, its maximum is 50, and its interface step is 2. A larger window gives the method a wider area in which to find similarities, but increases processing time. OpenCV specifies an odd window size; the default and the interface’s step sequence produce odd values, so keep it odd when supplying a value through a connection. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai))
template_window
An INT specifying the width, in pixels, of the small patch compared when the method looks for similarities. Its default is 3, its minimum is 3, its maximum is 20, and its interface step is 2. Changing this alters the size of the image details considered during comparison, so test it against the scale of the noise and textures you want to preserve. OpenCV specifies an odd patch size; the default and the interface’s step sequence produce odd values, so keep it odd when supplying a value through a connection. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai))
OpenCV Denoise (Luma/Chroma) Output Parameters:
image
The denoised IMAGE result. It contains one processed image for each incoming image, in the same batch order and at the same width and height. The node converts the OpenCV result back to RGB and returns floating-point image values scaled to the usual 0 to 1 range. Because processing passes through 8-bit image data, the result does not retain finer-than-8-bit precision from the input. This is the image to preview or pass to later editing steps.
OpenCV Denoise (Luma/Chroma) Usage Tips:
- Begin with the node defaults and compare the result with the original at full size. Increase
h_luminancein small increments when brightness grain remains; stop when important edges or textures begin to look soft. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai)) - If the main problem is colored speckles, try increasing
h_colorfirst while leavingh_luminancenear its default. This lets you address color noise without unnecessarily strengthening brightness smoothing. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai)) - Keep
search_windowsmall when processing speed matters. Increase it only when the wider search visibly improves the result, because larger search windows take longer to process. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai)) - Try
template_windowat3first, then compare odd sizes such as5or7on a representative crop containing both noise and fine detail. The patch size changes what the method considers similar; there is no single best value for every image. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai)) - Supply RGB images with three color channels. This implementation does not add an alpha channel to its result, so handle transparency separately if your workflow depends on it.
OpenCV Denoise (Luma/Chroma) Common Errors and Solutions:
RuntimeError: stack expects a non-empty TensorList
- Explanation: The incoming image batch was empty, so the node had no processed images to combine into its output.
- Solution: Connect a source that supplies at least one image and check whether an earlier filtering or batch step produced an empty result.
cv2.error: ... Invalid number of channels in input image
- Explanation: OpenCV could not perform the node’s RGB-to-BGR conversion because an incoming image did not have the expected three color channels. The exact wording may vary by OpenCV version.
- Solution: Convert grayscale images to RGB before this node. For images with transparency, prepare a three-channel RGB image and manage the alpha channel separately.
cv2.error: ... templateWindowSize ... or ... searchWindowSize ...
- Explanation: OpenCV expects odd sizes for these windows. A connected value may bypass the odd-number sequence offered by the node’s interface.
- Solution: Set
template_windowandsearch_windowto odd integers within their listed ranges, then run the node again. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai))
cv2.error: ... fastNlMeansDenoisingColored ...
- Explanation: OpenCV rejected the image or denoising configuration; its detailed error text depends on the installed version. The underlying function expects an 8-bit, three-channel image, which this node attempts to create from each incoming RGB image. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai))
- Solution: Check that the incoming images have three channels and usable dimensions, then retry with the default window settings. If the error persists, inspect the full OpenCV message for the condition it identifies.
RuntimeError: ... out of memory ... or the process stops during denoising
- Explanation: Large images or batches can require substantial memory, and a larger search window increases processing time. The exact message depends on where processing runs out of resources. (docs.opencv.org (https://docs.opencv.org/doc/doxygen/html/d1/d79/group__photo__denoise.html?utm_source=openai))
- Solution: Try one smaller image first, reduce the batch size, and return
search_windowto its default of5. Increase image size or the window only after a smaller run succeeds.
