🧹 Advanced Image Denoiser:
AdvancedImageDenoiser reduces visible grain and color speckles while aiming to preserve edges and useful texture. You can choose automatic noise-adaptive denoising or a method suited to a particular image, such as median filtering for isolated bright and dark specks. After denoising, it can selectively restore fine detail near edges, mix some of the original image back in, and apply optional sharpening. This gives you several ways to balance a clean result against the softness that stronger noise removal can introduce. It processes each image in a batch separately.
🧹 Advanced Image Denoiser Input Parameters:
image
The IMAGE to denoise; this required input has no numeric range or default. Connect an image or image batch. Each image is processed independently, so images with different noise levels can be handled in the same batch.
method
Selects the denoising approach. This required setting defaults to smart_auto; its options are smart_auto, non_local_means, bilateral, guided_filter, wavelet, total_variation, bm3d, and median. smart_auto measures noise and adjusts Non-Local Means denoising accordingly. non_local_means uses separate manual brightness and color-noise settings; bilateral and guided_filter also use those separate settings and aim to smooth while retaining edges. wavelet uses wavelet-based noise removal, total_variation favors piecewise-smooth results, bm3d uses the optional BM3D package, and median targets isolated impulse noise such as salt-and-pepper specks. If scikit-image is unavailable, wavelet and total_variation use the guided-filter fallback; if BM3D is unavailable, bm3d uses adaptive Non-Local Means instead.
strength
Controls denoising intensity for smart_auto, wavelet, total_variation, bm3d, and median. This required FLOAT defaults to 0.15 and ranges from 0.0 to 1.0 in 0.01 steps. Start low to protect texture: for smart_auto, 0.5 applies the measured-noise heuristic, while higher values can over-smooth. This setting does not control denoising intensity for the three manual brightness-and-color methods; use their separate strength controls instead. For wavelet, values at or above 0.5 use the full wavelet result rather than increasing its blend further.
detail_recovery
Controls how much fine detail from the original is restored after denoising. This required FLOAT defaults to 0.35 and ranges from 0.0 to 1.0 in 0.01 steps. The node uses edges detected in the cleaned image to favor detail restoration around edges and textured regions while leaving flatter areas cleaner. Raise it if denoising removes desirable texture, or lower it if grain becomes visible again near edges; values at or below 0.005 skip this step.
luminance_strength
Controls brightness-channel denoising for the manual non_local_means, bilateral, and guided_filter methods. This optional FLOAT defaults to 0.10 and ranges from 0.0 to 1.0 in 0.01 steps. Increasing it can clean brightness grain, but too much can soften fine shapes and texture. It does not set the brightness-denoising level for smart_auto or the other methods.
chroma_strength
Controls color-noise removal for the manual non_local_means, bilateral, and guided_filter methods. This optional FLOAT defaults to 0.30 and ranges from 0.0 to 1.0 in 0.01 steps. Increase it to reduce colored speckles or blotches; it can often be set higher than luminance_strength because smoothing color is generally less noticeable than smoothing brightness detail. It does not control color denoising for the other methods.
patch_size
Sets the comparison-patch size used by smart_auto and non_local_means. This optional INT defaults to 7, with a displayed range of 3 to 15 in steps of 2; the node uses an odd size. A patch is a small area the method compares with other areas to identify noise. The default is the suggested starting point; this setting has no effect on methods that do not use Non-Local Means.
search_window
Sets the search area used by smart_auto and non_local_means when looking for similar patches. This optional INT defaults to 21, with a displayed range of 7 to 35 in steps of 2; the node uses an odd size. A larger area gives the method more places to look for matches but takes longer to process. It has no effect on methods that do not use Non-Local Means.
wavelet_level
Sets the number of decomposition levels for wavelet denoising. This optional INT defaults to 3 and ranges from 1 to 6 in steps of 1. It changes the scales of detail considered by the wavelet method; it does not affect other methods, including the guided-filter fallback used when scikit-image is unavailable.
blend_original
Mixes the untouched image back into the denoised result after detail recovery. This optional FLOAT defaults to 0.0 and ranges from 0.0 to 1.0 in 0.01 steps. A small amount, such as 0.10–0.20, can make the result look more natural, but it also brings back some noise. At 1.0, the blend restores the original image before any optional sharpening; values at or below 0.005 skip blending.
sharpen_mode
Selects optional sharpening after denoising, detail recovery, and original-image blending. This optional setting defaults to off; its options are off, unsharp_mask, and luminance_only. unsharp_mask sharpens the image directly, while luminance_only sharpens its brightness channel without directly sharpening color channels, which helps avoid color fringing. Sharpening is applied only when the mode is not off and sharpen_amount is above 0.005.
sharpen_amount
Sets the intensity of optional post-sharpening. This optional FLOAT defaults to 0.20 and ranges from 0.0 to 1.0 in 0.01 steps. Increase it cautiously if the cleaned result needs more definition; too much can emphasize remaining noise or create harsh-looking edges. It has no effect when sharpen_mode is off, and values at or below 0.005 skip sharpening.
sharpen_radius
Controls the detail scale affected by optional post-sharpening. This optional FLOAT defaults to 0.30 and ranges from 0.05 to 1.0 in 0.01 steps. Lower values focus on finer detail; higher values broaden the sharpening effect. It has no visible effect when sharpening is disabled.
🧹 Advanced Image Denoiser Output Parameters:
image
The processed IMAGE, with one result for each input image in the same batch order. It contains the selected denoising result followed, when enabled, by detail recovery, original-image blending, and sharpening. Its appearance depends on those settings and on whether an optional method uses its fallback.
noise_report
A STRING with one line per processed image. Each line identifies the zero-based image index, the estimated noise standard deviation on a 0–255 scale, a low, moderate, or high label, and the selected method name. For example, img 0: noise sigma=5.00/255 (moderate), method=smart_auto. The labels mean low for a value below 3, moderate for a value from 3 up to but not including 8, and high for 8 or above. The report names the selected method even when a missing optional package causes that method to use a fallback.
🧹 Advanced Image Denoiser Usage Tips:
- Start with
smart_auto,strength=0.15, anddetail_recovery=0.35. Check the image before raising strength: stronger denoising may remove texture along with noise. - If the main problem is colored speckles, try
non_local_means,bilateral, orguided_filterwithchroma_strengthhigher thanluminance_strength; the defaults of0.30and0.10provide a starting point. - For isolated bright and dark pixels, try
median. For illustrated or cartoon-like images that should retain broad, smooth regions, trytotal_variation. - If edges look too soft, adjust
detail_recoverybefore adding sharpening. If the result then looks too processed, tryblend_originalaround0.10–0.20, keeping in mind that this also restores noise. - Keep
sharpen_modeatoffuntil you have evaluated the denoised result. If sharpening is needed, tryluminance_onlywith a modestsharpen_amountto avoid directly sharpening color noise. - For
smart_autoornon_local_means, keeppatch_size=7andsearch_window=21initially. Increase the search window only if the improvement justifies the longer processing time. - Read
noise_reportfor each image in a batch rather than assuming every image started with the same amount of noise. Its noise estimate describes the input image, not a measurement of noise remaining after processing.
🧹 Advanced Image Denoiser Common Errors and Solutions:
[AdvancedDenoiser] scikit-image missing — using guided filter
- Explanation: You selected
waveletortotal_variation, but scikit-image is unavailable. The node prints this message and produces a guided-filter result instead. - Solution: If you want the selected method, install scikit-image in the environment running ComfyUI and restart it. Otherwise, choose
guided_filterand adjust the settings for that result.
[AdvancedDenoiser] bm3d not installed — using adaptive NLM (pip install bm3d to enable)
- Explanation: You selected
bm3d, but its optional package is unavailable. The node continues with adaptive Non-Local Means denoising rather than BM3D. - Solution: Install
bm3din the environment running ComfyUI and restart it if you need BM3D; otherwise, usesmart_autofor an explicitly selected adaptive Non-Local Means workflow.
Output looks blurred or lacks texture
- Explanation: Denoising may be too strong for the image, especially when brightness detail is heavily smoothed. This is a visual result, not an error message.
- Solution: Lower
strengthfor applicable methods, or lowerluminance_strengthfor manual methods. Then try a modest increase indetail_recoverybefore adding sharpening.
Noise or colored speckles remain visible
- Explanation: The chosen method or strength may be too gentle for the noise. This is a visual result, not an error message.
- Solution: Increase
strengthgradually forsmart_autoor another strength-controlled method. With a manual method, raisechroma_strengthfor colored speckles orluminance_strengthfor brightness grain. Reduceblend_originalif it is bringing noise back.
Noise reappears around edges or after sharpening
- Explanation: Detail recovery can restore some original high-frequency information near edges, and sharpening can emphasize noise that remains. This is a visual result, not an error message.
- Solution: Lower
detail_recoveryorsharpen_amount, switch sharpening off, and check whetherblend_originalis also restoring noise.
Changing a setting appears to do nothing
- Explanation: Some controls apply only to particular methods: manual luminance and chroma strengths do not control
smart_auto;wavelet_levelapplies only towavelet; and sharpening controls have no effect whilesharpen_modeisoff. This is not an error message. - Solution: Check which method is selected and adjust its relevant controls. For sharpening, select a sharpening mode and set
sharpen_amountabove0.005.
Processing is slower than expected
- Explanation: A larger Non-Local Means search area increases the work per image, and some methods can be time-consuming, particularly on large images or batches. This is not an error message.
- Solution: Return
search_windowto its default of21if you increased it, test settings on a smaller image or batch, and compare methods to find an acceptable speed and quality balance.
Saved workflow settings produce unexpected results
- Explanation: Older workflows can restore widget values into different positions. The node accepts and clamps out-of-range values rather than rejecting the prompt, so the effective settings may differ from what you intended. This may occur without an error message.
- Solution: Recheck the selected method and every visible setting in an older workflow, then set them deliberately before processing.
