comfyui-advanced-denoiser Introduction
comfyui-advanced-denoiser is a ComfyUI custom node for cleaning up noisy images while preserving sharp edges and fine detail. It can help when grain, colored speckles, or other unwanted noise distracts from an otherwise good image—but ordinary smoothing makes the result look soft.
The node offers several denoising methods for different kinds of images. Its smart_auto method measures the image’s noise and adjusts the cleanup accordingly. After denoising, an optional edge-aware detail recovery pass can bring back texture near visible edges without restoring as much noise in smooth areas.
You can find the node under image → denoising in ComfyUI.
How comfyui-advanced-denoiser Works
Think of denoising as deciding which tiny variations belong in the artwork and which are unwanted speckles. Applying too much cleanup removes both. This node gives you ways to make that trade-off more carefully:
- Measure the noise. The node estimates how much noise is present. In
smart_auto, that measurement helps set the denoising level rather than relying entirely on a fixed guess. - Clean the image. The selected method reduces noise. Some methods are better suited to photo-like grain; others are useful for flat colors, strong edges, or isolated specks.
- Recover detail where it matters.
detail_recoveryuses edges detected in the cleaned image to guide the return of fine detail from the original. Smooth regions remain cleaner, while features such as outlines and textured edges can retain more definition. - Adjust the final appearance. You can mix a little of the untouched image back in or apply optional sharpening.
The noise_report output provides the measured noise level, called sigma, for each image. If you connect it to a text display node, you can see what the node detected while comparing settings.
comfyui-advanced-denoiser Features
A choice of denoising methods
| Method | When to try it | What to expect |
|---|---|---|
| smart_auto | Your first choice for most images | Measures noise and automatically adjusts non-local-means denoising. |
| non_local_means | Photo-like grain, especially when you want separate control over color noise | Lets you adjust luminance and chroma cleanup independently. |
| bilateral | Portraits or images with distinct edges | Smooths noise while working to preserve boundaries. |
| guided_filter | When you want fast, edge-preserving smoothing | Provides a lightweight smoothing option. |
| wavelet | Fine, evenly distributed grain | Uses BayesShrink with an automatically estimated noise level for each channel. |
| total_variation | Flat, synthetic, or AI-generated images | Can smooth broad areas strongly while retaining edges. |
| bm3d | When you want to try a quality-focused method and do not mind a slower run | Requires the optional bm3d package; otherwise, the node falls back to adaptive non-local means. |
| median | Isolated bright or dark specks, often called salt-and-pepper noise | Targets impulse noise rather than general photo grain. |
These are processing methods, not separate AI models. You do not need to choose a different image-generation model to use them.
Controls for balancing cleanliness and detail
strengthcontrols the amount of denoising. The author recommends starting low—around 0.1–0.25—because higher values can oversmooth an image. The default is 0.15. Insmart_auto, 0.5 applies the measured noise level, while 1.0 applies twice that level.detail_recoverybrings back fine texture near edges after cleanup. Try 0.3–0.6 if a result looks too soft. Because recovery is guided by edges in the denoised image and the measured noise level, it is designed to avoid bringing noise back into flat areas.luminance_strengthandchroma_strengthprovide separate control in applicable manual methods. Luminance affects brightness detail, where too much smoothing is especially noticeable. Chroma affects color variation; if you see colored speckles, you can often smooth chroma more strongly—around 2–3 times the luminance amount—without making the image look as blurry.blend_originalmixes some untouched input back into the finished image. A value around 0.1–0.2 can give the result a more natural look, though it can also return some noise.sharpen_modeoffers optional sharpening after denoising. Theluminance_onlyoption sharpens brightness detail without sharpening color differences, helping avoid colored fringes around edges.
A focused ComfyUI panel
The node includes a slate-and-teal UI panel with a method picker, short descriptions, and sliders that change to show parameters relevant to your selected method. Less-used controls sit in a collapsible advanced section. The panel’s values stay synchronized with standard ComfyUI widgets, so saved workflows and API use continue to work.
A practical starting workflow
- Select
smart_autoand setstrengthbetween 0.1 and 0.25. - Inspect both detailed areas and smooth backgrounds. If the image looks too soft, try
detail_recoverybetween 0.3 and 0.6. - If the remaining noise is mostly colored speckles, try
non_local_meansand increasechroma_strength. - If you want a less processed finish, try
blend_originalaround 0.1–0.2. - Connect
noise_reportto a text display node if you want to see the measured noise level for each image.
Troubleshooting comfyui-advanced-denoiser
The image looks blurry or overly smooth. What should you do?
- Lower
strength; forsmart_auto, return to the suggested 0.1–0.25 range. - Increase
detail_recoverygradually and compare important edges and textures. - Try a small amount of
blend_originalif you prefer a more natural finish.
Colored speckles remain after denoising. What should you do?
- Switch to
non_local_meansfor separate brightness and color-noise controls. - Increase
chroma_strengthwhile keeping luminance smoothing modest. - Check whether the color noise is reduced without sacrificing details you want to keep.
The noise consists of isolated bright or dark dots. Which method fits?
Try median. It is intended for salt-and-pepper artifacts, rather than fine grain spread across the whole image.
Why does selecting bm3d give an unexpected result?
The bm3d method needs the optional bm3d package. If it is unavailable, the node falls back to adaptive non-local means. If you cannot use that package, choose another method directly so you know which processing approach you are comparing.
Where can you see the detected noise level?
Connect the noise_report STRING output to a ComfyUI text display node. It reports the measured noise sigma for each image. You do not need to read this value to use smart_auto, but it can be useful when comparing images and settings.
