-vf hqdn3d=8:6:12:9 cleans grain out of 720p video in a few seconds and halves the file: 3.5 MB against 6.9 MB for a plain re-encode. ReelSDK's API took about 17 seconds for that job, under three quarters of a cent at $0.02 per compute minute.

Denoise video with ffmpeg

Grain costs bits, so a denoiser is also a compressor, and ReelSDK bills by the second, so speed counts too. Of 4 filters on 720p video, bm3d scored best on PSNR but took about 20 times as long as hqdn3d.

GuidesUpdated

hqdn3d
ffmpeg -i noisy.mp4 -vf hqdn3d=8:6:12:9 -c:v libx264 -crf 20 -preset veryfast -an denoised.mp4

Four strengths: spatial luma and chroma, then temporal luma and chroma. The defaults are mild; 8:6:12:9 took the grain out of the sky in a few seconds and wrote about half the bytes of a plain re-encode.

Variations

The other commands you will need

  1. Best detail: bm3d

    Block matching finds similar patches and filters them together. It scored highest against the clean copy and kept the stonework sharp, but took over a minute for 5 seconds of 720p.

    bm3d
    ffmpeg -i noisy.mp4 -vf bm3d=sigma=10 -c:v libx264 -crf 20 -preset veryfast -an bm3d.mp4
  2. Smoothest: nlmeans

    Non-local means compares each patch with its neighbours. At s=5 it made the smallest file, about a third of a plain re-encode, but smeared fine texture and took over 2 minutes, the slowest run here.

    nlmeans
    ffmpeg -i noisy.mp4 -vf nlmeans=s=5 -c:v libx264 -crf 20 -preset veryfast -an nl5.mp4
  3. Hiss in the audio

    afftdn removes steady background noise from the sound. nf sets the noise floor in dB, from -80 to -20. Copying the video stream kept the run under a second.

    audio
    ffmpeg -i in.mp4 -af afftdn=nf=-25 -c:v copy -c:a aac -movflags +faststart quiet.mp4

The flags

Four filters on the same noisy clip

PSNR against the clean copy, higher is closer. Every output was encoded with libx264 at CRF 20.

Four filters on the same noisy clip
FlagWhat it doesUsed here
hqdn3d=8:6:12:9Fast spatial and temporal smoothing36.5 dB, 3.5 MB, 3.2 s
hqdn3dThe same filter at its mild defaults36.9 dB, 5.8 MB, 4 s
atadenoiseAverages each pixel over 9 frames by default36.7 dB, 6.6 MB, 4.2 s
bm3d=sigma=10Filters groups of matching blocks together37.4 dB, 4.5 MB, 64 s
nlmeansNon-local means at strength 1, the default36.6 dB, 6.8 MB, 120 s
nlmeans=s=5The same at strength 5, of 1 to 3036.3 dB, 2.4 MB, 157 s
no filterPlain re-encode, for comparison36.6 dB, 6.9 MB, 4.6 s
noisy inputnoise=alls=20:allf=t+u, CRF 2336.1 dB, 6.4 MB

Source: ffmpeg filters: hqdn3d, nlmeans, atadenoise, bm3d, checked November 2026.

What the numbers say

PSNR moved less than you might expect: under a point and a half from the noisy input to bm3d, the best score. Even the plain re-encode gained about half a point, because x264 at CRF 20 drops some grain on its own. Only bm3d clearly beat it.

The bigger change was size. Random grain is the hardest thing for an encoder to compress, so taking it out first let hqdn3d at 8:6:12:9 hit about half the bytes of the plain re-encode at a near-identical score. In our side-by-side crops its sky came out clean.

Picking one

For most clips, start with hqdn3d: it is the fastest here and the four numbers are easy to tune. With no options, luma_spatial is 4.0 and the other three strengths follow from it. Raise the last two, the temporal ones, for grain that changes every frame.

Use bm3d when detail matters and you can wait, nlmeans when you want a smooth, almost painted look, and atadenoise for a quick temporal pass.

Running it on the API

All of these filters, and afftdn for sound, run on POST /v1/ffmpeg with the same args. Slow filters need care there: the hqdn3d job took about 17 seconds on the API against about 3 on our desktop, so a filter that took over 2 minutes locally can run into the 600-second job limit. Test on a short cut first and set timeout_s.

See pricing for the per-minute rate.

Without installing ffmpeg

The same command as one API call

Send the hqdn3d command's args to POST /v1/ffmpeg; the noisy clip is the one we made for this page. Each job runs ffmpeg 8.1 on a fresh machine at $0.02 per compute minute, billed per second; this one finished in about 17 seconds and cost under three quarters of a cent.

terminalTry it
curl "https://api.reelsdk.com/v1/ffmpeg?wait=60" \
  -H "Authorization: Bearer $REELSDK_KEY" \
  --json '{
    "inputs": {
      "noisy.mp4": "https://api.reelsdk.com/f/lib/in/guides/ffmpeg-denoise/noisy.mp4"
    },
    "args": [
      "-i", "noisy.mp4", "-vf", "hqdn3d=8:6:12:9", "-c:v", "libx264", "-crf",
      "20", "-preset", "veryfast", "-an", "denoised.mp4"
    ],
    "outputs": ["denoised.mp4"]
  }'
denoised.mp4 from the API: 5 seconds of 720p at about 3.5 MB.Input: noisy.mp4

FAQ

Questions, answered

What is the best ffmpeg filter to remove video noise?

bm3d=sigma=10 scored best against the clean copy but took over a minute for 5 seconds of 720p. hqdn3d=8:6:12:9 took a few seconds and halved the file size.

Does denoising make the file smaller?

Usually, yes. Grain is hard to compress: hqdn3d=8:6:12:9 wrote about 3.5 MB where a plain re-encode wrote 6.9 MB at the same CRF.

Why is nlmeans so slow?

It compares each patch with every patch in its research window, 15 by default. On 5 seconds of 720p it took about 2 minutes at the default strength and longer at s=5.

How do I remove hiss from the audio?

Use -af afftdn with -c:v copy so the video is not re-encoded. nf sets the noise floor, from -80 to -20 dB; the default is -50.

Ready to build?

Ship video from your app today

Your first $0.50 is free, and one key covers every endpoint.

Moonlit lake with a lit cabin under the mountains