Guide
Why Denoise Video Before Encoding?
Noise is random. An encoder cannot predict it, so it spends bits on it that never reach the viewer as real picture information. On real content, noise and grain alone take up to 30% of the bitrate budget. This guide explains where these bits go, why block-based codecs suffer most, and how to remove noise without losing detail.
The short answer
- Noise has no temporal redundancy, so inter prediction fails and the encoder codes it as new information in every frame.
- Noise is mistaken for motion, which leads to faulty motion vectors and higher bitrates.
- Removing noise before encoding frees the encoder for real image content and enables up to 30% lower bitrates at the same quality.
- The denoiser must separate noise from detail. A filter that blurs saves bits by destroying the picture.
Noise breaks motion vectors
Modern encoders save most of their bits through temporal prediction: a block in the current frame is described as a shifted block from a previous frame plus a small residual. This works because real image content moves in a predictable way. Noise does not. Every frame carries a new random pattern, so the residual stays large no matter how good the prediction is.
Worse, the motion search itself is disturbed. Noise can be mistaken for motion, the encoder picks faulty motion vectors, and the residual grows further. In flat and dark areas, where the real signal is weak compared to the noise amplitude, this effect is strongest. That is exactly where HDR content shows more detail than ever and where viewers notice artifacts first.
Why block-based codecs suffer
H.264, HEVC and VVC split the picture into blocks, transform each block into frequencies and quantize them. The high frequencies, where noise and grain live, are quantized hardest. The result is well known: smooth, plastic areas at best, blocky areas and banding at worst.
Noise also amplifies the distortions between blocks. Low-frequency components span larger areas and are sensitive to block boundaries, so noise makes the edges of the blocks visible. Denoising before encoding mitigates this effect for any block-based codec. For film content with intended grain, the grain can be removed before encoding and added back on playback with AV1 film grain synthesis.
Noise in a statistical multiplex
In broadcast, several services often share one transport stream with a fixed total bitrate. The multiplexer gives more bits to the service that currently needs them. A noisy service always looks complex, so it takes bits away from its neighbours. One noisy channel can lower the picture quality of several others. Cleaning the noisy source raises the quality of the whole multiplex.
Denoising without losing detail
The hard part is not removing noise, it is keeping everything else. Image information is part of the video noise floor: fine textures, skin detail and film structure sit at the same amplitude as the noise. A classic spatial or temporal filter cannot tell them apart and smooths both. The bitrate drops, but so does the resolution, and the saving is paid for with detail.
IRIS takes a different approach: it separates noise from real image content and lifts the signal out of the noise floor. Perceived resolution goes up instead of down. This is also why the term "visually lossless" is misleading: even compression ratios below 10:1 cost image information. The less noise the encoder has to carry, the more of the real picture survives.
Where in the chain to denoise
Noise management does not need to run at every stage. There are three effective entry points:
- Source driven: denoise RAW data before debayering and before the creative pipeline with IRIS.RAW and BRAW2IRIS. This gives the best signal-to-noise ratio for grading and compositing.
- Master driven: clean an intermediate such as FFV1 or ProRes after grading and VFX.
- Distribution driven: denoise per shot right before the encoder, combined with per-shot encoding, or in real time for live production with IRIS.Broadcast.
A signal produced in 4K can be denoised once and then downscaled for every lower resolution. Second screens have no built-in denoising or deblocking, so optimizing the video at the source improves every simulcast feed.
How to measure the result
Standard metrics compare an encode against its source and count every difference as an error, including removed noise. Measured against the noisy original, a denoised encode drops into the 30 dB range in PSNR simply because the removed noise counts as loss. VMAF, on the other hand, is largely blind to high-frequency noise and scores a denoised file against the original between 98 and 100.
For a fair comparison, use the denoised source as the reference for the denoised workflow and the original as the reference for the standard workflow. Then compare quality stability, not only the average, for example with QBVE. In a critical low-light scene of Netflix "Meridian" at 9.6 Mbit/s UHD HEVC, denoising with IRIS reduced the frames below VMAF 93 from 25.73% to 13.29% and the frames below 40 dB PSNR from 51.62% to 10.15%. More on metrics in our guide VMAF mean, harmonic mean, 1% low and QBVE.
Try it on your own material
We run a free comparison on a sample of your content: standard encode against IRIS-Denoising at the same settings, with VMAF, PSNR and QBVE. Contact info@waveletbeam.com.