Remove Image Noise with a Median Filter
Pick an image and the median filter replaces every pixel with the middle value of its neighbours — isolated dots and dust disappear, edges stay crisp. Nothing is uploaded.
What a median filter does
A median filter slides a small square window over the image and replaces the centre pixel with the middle value of everything inside that window. Because the middle value is a real pixel from the neighbourhood — not an average — a lone white or black dot has no influence at all: it sorts to the end of the list and is thrown away, while the edge running through the window keeps its two populations of pixels and stays exactly where it was.
That is why median filtering is the standard answer to salt-and-pepper noise (scattered black/white dots from bad sensor pixels, transmission errors or a dusty scan) and why it beats a Gaussian blur, which smears the dot into a grey smudge and softens every edge along with it.
Worked example
Take a 3-pixel-wide row of a photo where one pixel has been hit by noise:
neighbours: 118 120 255 121 119
sorted: 118 119 120 121 255
median: 120
The blown-out 255 is discarded and the pixel becomes 120, matching its
neighbours. An average would have produced 146 — a visible bright smudge.
On this page: upload a noisy image, leave Window radius at 1 (a 3×3
window) and press run. Single-pixel dots are gone and the result downloads in
the same format you uploaded. Raise the radius to 2 (5×5) for scanner dust or
speckle that covers a couple of pixels.
The controls
- Window radius — the window is
2 × radius + 1square.1= 3×3,2= 5×5,5= 11×11,20= 41×41. Bigger windows erase bigger blobs of noise but also erase fine texture, and the filter gets quadratically slower. - Specks to remove — both is the true median. Bright only leans toward the darker neighbours, so white dust is wiped harder (the image darkens a little); dark only does the reverse for black specks on light paper.
- Channels to filter — all cleans everything. Luma only smooths brightness noise while colour detail is untouched. Chroma only removes the blotchy red/green patches of high-ISO colour noise and leaves luminance detail pin-sharp — usually the best-looking option for photos.
- Passes — running a small window twice removes dense noise more gently than one large window, because each pass only ever moves a pixel to a value that already exists nearby.
- Output format / quality — keep returns the format you uploaded; PNG is lossless; JPG and WebP use the quality slider (1–100, default 92).
- Strip metadata — drops EXIF, camera info and comments from the output.
Limits and edge cases
- Input files up to 8 MB, any format your browser's decoder handles (PNG, JPG, WebP, GIF, BMP, TIFF).
- Radius is capped at 20 (a 41×41 window) and passes at 3. Large radius values on a multi-megapixel photo can take a while — the work grows with the square of the window.
- The median filter removes impulse noise. Fine film grain or heavy Gaussian sensor noise is better handled by a dedicated grain/denoise pass; a median filter will just flatten texture.
- Text, thin lines and small dots are also "isolated pixels": at radius 3 and above, fine line art and small type start to erode. Keep the window as small as it can be while still catching the noise.
- Luma only and chroma only convert the image through Y'CbCr, which can shift untouched channels by ±1 sRGB step. All filters the image in its native planar layout, with no colour conversion.
- Choosing PNG, JPG or WebP for an animated GIF keeps only the first frame; keep filters every frame and stays animated.
FAQ
What is the difference between a median filter and a blur?
A blur averages the pixels in the window, so a bright noise dot contaminates every pixel around it and edges soften. A median selects a value that already exists in the window, so an outlier is discarded outright and a straight edge stays straight. For salt-and-pepper noise the median wins clearly; for smooth grain, a blur or a dedicated denoiser is a better fit.
What radius should I use?
Start at 1 (a 3×3 window) — it removes single-pixel dots with almost no loss
of detail. Use 2 (5×5) if the specks are 2–3 pixels wide, such as dust on a
scan. Above 4 the image starts to look like a watercolour, so if you need that
much cleaning, try 2 passes at a small radius instead of one big window.
Why does the result look flat or "painted"?
Because the window is too large for the detail in the image. Every pass replaces texture with the local middle value, so fine grain, skin texture and small type get flattened. Lower the radius, drop to a single pass, or switch Channels to filter to chroma only so luminance detail is left completely untouched.
Can it clean up a scanned document?
Yes — that is the classic use. A radius of 1–2 with dark only removes pepper specks from light paper without eating the strokes of the text, and bright only removes white pinholes from a dark background. Choose PNG output so the cleaned scan is not re-compressed, and tick Strip metadata if you are about to publish it.
Are my images uploaded anywhere?
No. The filter runs in your browser through a WebAssembly build of ffmpeg. The file you pick is read locally, processed locally, and the download link points at the result held in your own tab's memory.
Developer & Automation Access
Run it from the terminal
Same engine as this page, headless — via the gizza CLI:
gizza tool image-median-denoise 'url=https://example.com/input' 'radius=1' 'target=both' 'channels=all' 'passes=1' 'format=keep' 'quality=92' 'strip_metadata=true'New to the CLI? Get gizza →
Open it by URL
Pre-fill and auto-run this tool with query parameters — the names match the API/CLI:
https://gizza.ai/tools/image-median-denoise/?url=https://example.com/input&radius=1&target=both&channels=all&passes=1&format=keep&quality=92&strip_metadata=trueMachine-readable descriptor: tool.json — title + parameters JSON Schema for agents.