Video Denoising – Complete Guide to Noise Removal

Noise is the enemy of clean video. Whether it's the grain from a high‑ISO shot, the speckles from an old analog capture, or the blocky artifacts from a low‑bitrate encode, noise degrades quality, wastes bitrate, and makes your footage look amateurish.

FFmpeg has over a dozen dedicated denoise filters[reference:0]—from the fast and reliable hqdn3d to the cutting‑edge nlmeans and bm3d. But with so many options, how do you choose? The documentation lists parameters but doesn't tell you which filter to use or how to tune it.

This guide changes that. You'll learn about every major denoise filter, when to use each one, and how to get professional results.

The Gap: A Dozen Filters, Almost No Guidance

FFmpeg supports multiple denoise filters[reference:4], but the documentation is sparse. The Denoise Examples page lists filters and basic commands but doesn't explain why you'd choose one over another. As the page itself admits: "This document does not offer direct recommendations for denoise video algorithm selection, since the expected quality and performance is entirely subjective and all sources differ"[reference:5].

This guide fills that gap with practical advice, real‑world examples, and performance comparisons.

The Symptom: "My Video Is Too Noisy"

You've shot footage with a smartphone at high ISO, or digitized an old VHS tape, or received a low‑bitrate MP4 from a client. The video is grainy, speckled, or blocky. You need to clean it up—but you don't want to blur away all the detail.

This is the classic denoising dilemma: remove noise without destroying detail. FFmpeg's filters give you the tools to find the right balance.

Core Concepts: Understanding Noise

Before diving into filters, it helps to understand what you're dealing with.

Types of video noise:

• Gaussian noise — random variations in brightness and color. Common in low‑light footage.
• Salt‑and‑pepper noise — random white and black pixels. Common in old digitizations.
• Compression artifacts — blocky, mosquito noise from low‑bitrate encoding.
• Film grain — natural texture from analog film (sometimes desirable).
• Chroma noise — color noise in the chrominance channels.

Denoising approaches:

• Spatial denoising — looks at individual frames, averaging neighboring pixels. Good for static noise but can blur edges.
• Temporal denoising — looks across multiple frames, averaging over time. Excellent for removing random noise without blurring[reference:6].
• Frequency‑domain denoising — transforms the image to the frequency domain (FFT, DCT, wavelet) and filters out high‑frequency noise[reference:7].
• Patch‑based denoising — finds similar patches across the image (nlmeans, bm3d) for state‑of‑the‑art results[reference:8].

Most FFmpeg denoise filters combine multiple approaches.

Denoise Filters at a Glance

Here's a quick overview of the major denoise filters in FFmpeg[reference:9]:

hqdn3d – The Workhorse

hqdn3d (High Quality 3D Denoiser) is the most popular denoise filter in FFmpeg[reference:20]. It combines spatial and temporal denoising for excellent results at reasonable speed.

Syntax:

hqdn3d=luma_spatial=4.0:chroma_spatial=3.0:luma_tmp=6.0:chroma_tmp=3.0

Parameters:

• luma_spatial — Spatial denoising strength for luma (brightness). Default 4.0.[reference:21]
• chroma_spatial — Spatial denoising strength for chroma (color). Default 3.0.[reference:22]
• luma_tmp — Temporal denoising strength for luma. Default 6.0.[reference:23]
• chroma_tmp — Temporal denoising strength for chroma. Default 3.0.[reference:24]

Example:

# Basic hqdn3d denoising with moderate settings
ffmpeg -i noisy.mp4 -vf "hqdn3d=4:3:6:4.5" -c:v libx264 -crf 20 denoised.mp4

This command applies moderate denoising[reference:25]. The parameters are luma_spatial:chroma_spatial:luma_tmp:chroma_tmp.

Tips:

• Start with 4:3:6:4.5 and adjust up or down.
• For very noisy footage, try 6:4:8:6.
• To denoise only luma (leaving color untouched), set chroma parameters to 0: hqdn3d=4:0:6:0.
• To denoise only chroma (color noise), set luma parameters to 0: hqdn3d=0:3:0:4.5.

nlmeans – The Quality Champion

nlmeans (Non‑Local Means) is widely considered one of the most advanced denoise filters[reference:26]. It works by finding similar patches across the image and averaging them—preserving edges and textures much better than simpler filters[reference:27].

Limitation: Currently limited to 8‑bit per channel formats[reference:28].

Syntax:

nlmeans=s=STRENGTH:r=PATCH_RADIUS:p=PREVIEW

Parameters:

• s — Denoising strength. Default 1.0, range 1.0–30.0.[reference:29]
• r — Patch radius. Default 3.
• p — Preview mode (0=off, 1=on).

Example:

# nlmeans denoising with strength 30 (aggressive)
ffmpeg -i noisy.mp4 -vf "nlmeans=s=30:r=3:p=1" -c:v libx264 -crf 20 denoised.mp4

As one user noted, nlmeans is "too excellent" and can recover even bad sources like VHS video[reference:30].

Tips:

• Start with s=4 for light denoising, s=10 for moderate, s=20-30 for aggressive[reference:31][reference:32].
• p=1 enables preview mode—useful for testing.
• For best results, pair with unsharp to restore sharpness: nlmeans=s=10:r=3,unsharp=5:5:1.0:5:5:0.0[reference:33].
• nlmeans is significantly slower than hqdn3d[reference:34].

vaguedenoiser – Wavelet Precision

vaguedenoiser applies a wavelet‑based denoiser using the Cohen‑Daubechies‑Feauveau 9/7 wavelet[reference:35]. It transforms each frame into the wavelet domain, filters coefficients, and inverse transforms[reference:36].

Note: Requires FFmpeg to be compiled with --enable-libvidstab[reference:37].

Syntax:

vaguedenoiser=threshold=THRESHOLD:method=METHOD:nsteps=NSTEPS:percent=PERCENT:planes=PLANES:type=TYPE

Parameters:

• threshold — Filtering strength. Higher = more filtering. Default 2.[reference:38]
• method — hard, soft, or garrote (default)[reference:39].
• nsteps — Wavelet decomposition levels. Default 6[reference:40].
• percent — Percentage of full denoising. Default 85[reference:41].
• planes — Which planes to process. Default all[reference:42].
• type — universal (default) or bayes[reference:43].

Example:

# vaguedenoiser with moderate threshold
ffmpeg -i noisy.mp4 -vf "vaguedenoiser=threshold=22:percent=100:nsteps=4" -c:v libx264 -crf 20 denoised.mp4

This applies wavelet denoising with a threshold of 22 and 4 decomposition levels[reference:44].

Tips:

• method=garrote (default) is a good balance between soft and hard thresholding[reference:45].
• nsteps values above 6 may not improve results much[reference:46].
• percent=100 applies full denoising; lower values apply partial[reference:47].
• type=bayes can be more adaptive than universal[reference:48].

bm3d – State of the Art

bm3d (Block‑Matching 3D) is one of the most advanced denoising algorithms available[reference:49]. It works by finding similar blocks in the image, grouping them into 3D stacks, applying collaborative filtering, and aggregating the results[reference:50].

Syntax:

bm3d=sigma=SIGMA:block=BLOCK:bstep=BSTEP:group=GROUP:range=RANGE:mstep=MSTEP:thmse=THMSE:hdthr=HDTHR:estim=ESTIM:planes=PLANES

Parameters:

• sigma — Denoising strength. Default 1, range 0–999.9.[reference:51]
• block — Block size. Default 4.[reference:52]
• bstep — Block step. Default 2.[reference:53]
• group — Group size. Default 1.[reference:54]
• estim — Estimation mode: basic or final[reference:55].

Example:

# bm3d denoising with sigma 30
ffmpeg -i noisy.mp4 -vf "bm3d=sigma=30:block=4:bstep=2:group=1:estim=basic" -c:v libx264 -crf 20 denoised.mp4

This applies BM3D denoising with a sigma of 30[reference:56]. BM3D is extremely slow but delivers state‑of‑the‑art quality[reference:57].

atadenoise – Fast & Adaptive

atadenoise (Adaptive Temporal Averaging Denoiser) is one of the fastest denoise filters[reference:58]. It works by averaging pixels across multiple frames, adapting to motion to avoid ghosting.

Syntax:

atadenoise=0a=THRESHOLD_A:0b=THRESHOLD_B:1a=THRESHOLD_A:1b=THRESHOLD_B:2a=THRESHOLD_A:2b=THRESHOLD_B

Parameters:

• 0a — Threshold A for plane 0 (luma)[reference:59]
• 0b — Threshold B for plane 0 (luma)[reference:60]
• 1a — Threshold A for plane 1 (chroma)
• 1b — Threshold B for plane 1 (chroma)

Example:

# atadenoise with default settings
ffmpeg -i noisy.mp4 -vf "atadenoise" -c:v libx264 -crf 20 denoised.mp4

With default parameters, atadenoise is a great starting point for fast denoising[reference:61].

fftdnoiz – Frequency Domain

fftdnoiz applies denoising using a 3D Fast Fourier Transform[reference:62]. "3D" means it considers both spatial and temporal dimensions[reference:63].

Syntax:

fftdnoiz=SIGMA:BLOCK:OVERLAP:PRECISION

Parameters:

• sigma — Denoising strength. Typical values 5‑9.[reference:64]
• block — Block size.
• overlap — Overlap between blocks.
• precision — Calculation precision.

Example:

# fftdnoiz with sigma 30
ffmpeg -i noisy.mp4 -vf "fftdnoiz=30:1:6:0.8" -c:v libx264 -crf 20 denoised.mp4

Warning: fftdnoiz can introduce ringing artifacts with bad settings[reference:65]. Start with lower sigma values (5‑9) and increase gradually[reference:66].

Other Denoise Filters

chromanr

Reduces chrominance (color) noise only. Useful when luma is clean but color is noisy[reference:67].

ffmpeg -i noisy.mp4 -vf "chromanr" -c:v libx264 -crf 20 denoised.mp4

removegrain

Spatial denoiser for progressive video[reference:68]. Fast but can blur edges.

ffmpeg -i noisy.mp4 -vf "removegrain=4" -c:v libx264 -crf 20 denoised.mp4

Users have reported success chaining multiple removegrain filters[reference:69].

tmedian

Temporal median filter—picks the median pixel from several frames[reference:70]. Good for heavy transient artifacts[reference:71].

ffmpeg -i noisy.mp4 -vf "tmedian=radius=3" -c:v libx264 -crf 20 denoised.mp4

tmix

Reduces noise by mixing successive video frames[reference:72].

ffmpeg -i noisy.mp4 -vf "tmix=frames=5" -c:v libx264 -crf 20 denoised.mp4

owdenoise

Overcomplete wavelet denoiser[reference:73]. Similar to vaguedenoiser but uses a different wavelet transform[reference:74].

ffmpeg -i noisy.mp4 -vf "owdenoise=8:6:6" -c:v libx264 -crf 20 denoised.mp4

dctdnoiz

2D DCT‑based denoiser[reference:75]. Extremely slow and not designed for real time[reference:76].

ffmpeg -i noisy.mp4 -vf "dctdnoiz=sigma=30:n=4" -c:v libx264 -crf 20 denoised.mp4

Practical Example: Denoising Pipeline

For the best results, combine denoising with sharpening:

# Professional denoising pipeline: denoise → sharpen
ffmpeg -i noisy.mp4 -vf "hqdn3d=4:3:6:4.5,unsharp=5:5:1.0:5:5:0.0" -c:v libx264 -crf 18 -preset slow denoised_sharp.mp4

For high‑quality denoising with nlmeans:

# High-quality denoising with nlmeans + unsharp
ffmpeg -i noisy.mp4 -vf "nlmeans=s=10:r=3:p=1,unsharp=5:5:1.0:5:5:0.0" -c:v libx264 -crf 18 -preset slow denoised_high_quality.mp4

Performance Comparison

Based on real‑world testing, here's how the filters compare for a 10‑second 1080p clip on a modern CPU[reference:77]:

As noted in the FFmpeg wiki, "these comparisons and values are not recommendations, but are to demonstrate how algorithm and options selection can have a major impact on performance"[reference:78].

Debugging Common Issues

Here are the most common denoising problems and how to fix them.

Pro tip: As one user noted, "I always use unsharp to sharpen the image after nlmeans"[reference:81]. Denoising always removes some detail—sharpening helps restore it.

Visualize Denoising with the FFmpegLab IDE

Denoising is much easier to tune with visual feedback. The FFmpegLab IDE lets you:

• Preview denoising in real time—adjust parameters and see the result instantly
• Compare before and after side by side
• Adjust strength sliders for hqdn3d, nlmeans, and vaguedenoiser
• Visualize noise reduction with a difference overlay
• Generate the exact command with your tuned parameters

Here's how denoising tuning looks in the FFmpegLab IDE.

# Visual denoising preview
# Adjust sliders → see results in real time

The IDE shows you exactly what's happening—the original frame, the denoised result, and metrics like noise reduction percentage. You can adjust parameters with sliders and see the results in real time.

Frequently Asked Questions (FAQ)

Final Word

Video denoising is one of FFmpeg's most powerful—and most complex—capabilities. With over a dozen dedicated filters, you have the tools to clean up almost any noisy footage.

The key is understanding the trade‑offs: hqdn3d for speed and reliability, nlmeans and bm3d for quality, vaguedenoiser for wavelet precision, and atadenoise for fast temporal denoising.

With the FFmpegLab IDE's visual feedback, you can experiment, debug, and perfect your denoising settings without the guesswork. Adjust parameters with sliders, preview the results in real time, and export the exact command you need.

Next time you're facing noisy footage, reach for FFmpeg's denoise filters. Your viewers will thank you.